Invasive Alien Plant Species Project UKZN/DAEA An investigation into using different satellite remote sensors and techniques to identify, map, monitor and predict the spread and distribution of some of the major current and emerging invasive alien plant species in KwaZulu-Natal Final Report (July 2006- January 2010) Submitted to: Invasive Alien Species Programme (DAEA)
Prepared by: Mr Innocent Z. Shezi & Mr Nitesh K. Poona
Supervisor: Prof Fethi Ahmed School of Environmental Sciences, University of KwaZulu-Natal, Howard College Campus, Durban
1
Table of Contents Table of Contents ............................................................................................................ 2 List of Figures ................................................................................................................. 5 List of Tables .................................................................................................................. 7 List of Abbreviations and Acronyms .............................................................................. 8 Executive Summary ........................................................................................................ 9 Chapter 1: General Introduciton ................................................................................... 13 1.1
Introduction ....................................................................................................... 13
1.2
IAP Species: Challenges and Opportunities ..................................................... 16
1.3
Remote Sensing, Geographic Information System and IAP Species Mapping 17
1.4
Motivation to the Study .................................................................................... 18
1.5
Aim of Study ..................................................................................................... 19
1.6
Objectives of Study ........................................................................................... 19
1.7
Description of Study Area ................................................................................ 20
1.8
IAP Species Selection ....................................................................................... 22
1.8.1 Acacia mearnsii (current invader): KwaZulu-Natal midlands...................... 24 1.8.2 Lantana camara (current invader): Coastal KwaZulu-Natal ........................ 24 1.8.3 Parthenium hysterophorus (emergent invader): Northern KwaZulu-Natal.. 25 1.8.4 Rubus cuneifolius (current invader): KwaZulu-Natal highlands .................. 26 1.9 SPOT 5 Satellite Imagery ................................................................................. 27 1.10
Organisation of the Report ................................................................................ 27
Chapter 2: Literature Review ........................................................................................ 29 2.1
Introduction ....................................................................................................... 29
2.2
IAP Species in Context: Impact on the Natural Resource Base ....................... 30
2.3
IAP Species in Context: Impact on Human Welfare and Quality of Life ........ 35
2.4
Detection and Mapping of IAP Species using Remote Sensing ....................... 37
2.5
Automated Procedures for Image Classification: Protocols and Algorithms for
Mapping IAP Species ................................................................................................... 38 2.6
Modelling IAP Species ..................................................................................... 40
2.7
Conclusions ....................................................................................................... 43
Chapter 3: Mapping Invasive Alien Plant Species ....................................................... 45 3.1
Introduction ....................................................................................................... 45 2
3.2
Materials and Methods ...................................................................................... 46
3.2.1 Study area and target species ........................................................................ 46 3.2.2 Data acquisition ............................................................................................ 47 3.2.3 Data processing and analysis ........................................................................ 49 3.2.3.1 Pre-processing ....................................................................................... 50 3.2.3.2 Whole-pixel training and classification ................................................ 51 3.2.3.3 Post-classification processing ............................................................... 55 3.3 Results and Discussion ..................................................................................... 57 3.3.1 Unsupervised classification .......................................................................... 57 3.3.2 Supervised classification ............................................................................... 59 3.3.2.1 Classification accuracy assessment: A. meanrsii distribution maps ......... 59 3.3.2.2 Classification accuracy assessment: R. cuneifolius distribution maps ..... 63 3.3.2.3 Classification accuracy assessment: P. hysterophorus distribution maps . 66 3.3.2.4 Classification accuracy assessment: L. camara distribution maps............ 68 3.3 3 Scale, Resolution, and IAP Species Mapping ................................................... 72 3.4
Conclusions ....................................................................................................... 72
Chapter 4: Semi-Automated Mapping .......................................................................... 74 4.1
Introduction ...................................................................................................... 74
4.2
Automating Image Processing: Platforms used to Create Protocols and
Algorithms .................................................................................................................... 75 4.3
Motivation, Aim and Objectives ....................................................................... 81
4.5
Material and Methods ...................................................................................... 82
4.5.1 Data collection: remotely sensed imagery and Global Positioning Systems data 82 4.5.2 Designing and automating an image analysis scheme: protocol and algorithm development .............................................................................................................. 82 4.6 Results and Discussions ................................................................................... 83 4.6.1 Automation in ArcGIS-ArcEditor 9.3: A semi-automatic algorithm for Mapping selected IAP species .................................................................................. 84 4.7 Conclusion ....................................................................................................... 89 Chapter 5: Modelling the Range and Distribution of IAPS .......................................... 90 5.1
Introduction ....................................................................................................... 90
5.2
Materials and Methods ...................................................................................... 92
5.2.1 Study area and target species ........................................................................ 92 5.2.2 Data acquisition ............................................................................................ 94 5.2.3 Data processing and analysis ........................................................................ 96 5.2.3.1 Mahalanobis distances .......................................................................... 96 5.2.3.2 Chi square P-values............................................................................... 97 5.2.4 Model assessment and validation.................................................................. 97 3
5.3
Results and Discussion ..................................................................................... 98
5.4
Conclusion ...................................................................................................... 101
Chapter 6: General Discussion ................................................................................... 103 6.1
Discussion and conclusions ............................................................................ 103
6.2
Recommendations ........................................................................................... 104
References: ....................................................................................................................................... 107 Appendix A: Systematic prioritisation matrix ................................................................................. 121 Appendix B: Statistical measures utilised in the study .................................................................... 123 Appendix C: Protocol and Algorithm DVD .................................................................................... 123 Appendix D: Detail A0 Maps of the Selected IAP Species Distribution Hardcopies and DVD ................................................................................................................................................. 123
4
List of Figures Figure 1.1: The study area of KwaZulu-Natal. Map (A) shows the four biomes of KwaZulu-Natal and the major fluvial systems draining the province, and map (B) indicates the four sites selected for the study lying within 4 vegetation types of KZN (after Low and Rebelo, 1998). .................................................................................. 21 Figure 1.2: The four IAP species selected for the study. Clockwise, from top left: Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius. ...... 23 Figure 2.1: The varied biomes of South Africa (map A) and KwaZulu-Natal (map B). . 33 Figure 3.1: Location of the study area and study sites. Map (A) shows the location of KwaZulu-Natal in South Africa, map (B) shows the four biomes of KwaZulu-Natal (after Low and Rebelo, 1998), map (C) shows the location of the four study sites, map (D) indicates the four SPOT 5 scenes covering the each of the four study sites, and map (E) indicates the scene acquired for the KZN-midlands. ........................... 47 Figure 3.2: The target species occurring in the field. Clockwise, from top left: Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius. ...... 49 Figure 3.3: Mean signature plot of 10-class ISODATA for SPOT 5. The target class is highlighted in green. ................................................................................................. 52 Figure 3.4: Unsupervised training of SPOT 5 using the ISODATA algorithm. Map (A) Acacia mearnsii, map (B) Lantana camara, map (C) Parthenium hysterophorus, and map (D) Rubus cuneifolius. Inserts show overlay with testing data (training sites). ................................................................................................................................... 58 Figure 3.5: Shows the distribution of A. mearnsii using ML, MDM, MAHAL and BOX classifier algorithms .................................................................................................. 62 Figure 3.6: Shows distribution of target species R. cunefolius in KZN highlands obtained using MAHAL, MDM, BOX and ML classifiers ..................................................... 65 Figure 3.7: Shows the distribution of P. hysterophorus in Northern KZN obtained using MAHAL, MDM, BOX and ML classifiers ............................................................... 67 Figure 3.8: Shows the distribution of L. camara in Coastal KZN obtained using MAHAL, MDM, BOX and ML classifiers ............................................................... 71 Figure 4.1: Portion of a script-based algorithm that runs via SML background. The algorithm converts digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map A. mearnsii..................... 77 Figure 4.2: A graphic based algorithm modelled in Model Maker. The algorithm converts digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map A. mearnsii. .................................. 78 Figure 4.3: Portion of a script-based based algorithm written in VBscript. The algorithm converts digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map Acacia. mearnsii. .......................... 79 Figure 4.4: A graphic-based algorithm modelled in ModelBuilder. The algorithm convert digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map Acacia. mearnsii. .......................... 79 Figure 4.5: is a schematic showing the basic components of used to build a model in ModelBuilder interface. Elements colours are used symbolise input data, geoprocessing tool and derived data is kept throughout the model. ......................... 85 Figure 4.6: A schematic representation of an algorithm generated and executed in ModelBuilder. Note colour code representing input elements, geoprocessing tools 5
and derived data. ....................................................................................................... 87 Figure 4.7: An application designed to run the algorithm ............................................... 88 Figure 5.1: Conceptual framework for the Mahalanobian model (adapted from Farber and Kadmon, 2003). .................................................................................................. 91 Figure 5.2: The bioclimatic regions of KwaZulu-Natal (after Phillips, 1973) indicating point locations identifying species‘ presence. Insert (a) Acacia mearnsii (KZN midlands), insert (b) Parthenium hysterophorus (northern KZN), insert (c) Rubus cuneifolius (KZN highlands), and insert (d) Lantana camara (coastal plains of KZN). ................................................................................................................................... 92 Figure 5.3: Habitat suitability using Mahalanobian model. Map (A) Acacia mearnsii, map (B) Lantana camara, map (C) Parthenium hysterphorus, and map (D) Rubus cuneifolius. Darker shading of pixels indicates increased habitat suitability. ....... 100 Figure A1: Systematic species‘ prioritisation matrix. .................................................... 121 Figure A2: Systematic species‘ prioritisation matrix key. ............................................. 122
6
List of Tables Table 1.1: Key research initiatives on alien plant invasions in South Africa (adapted from Richardson and van Wilgen, 2004). .......................................................................... 14 Table 2.1: Top 10 invading species in South Africa ranked by condensed invaded area (after Versveld et al., 1998). ..................................................................................... 31 Table 2.2: Most important invader species in KwaZulu-Natal ranked by condensed invaded area (after Versveld et al., 1998). The condensed area is the total area adjusted to bring the cover to the equivalent of 100%.............................................. 34 Table 2.3: Areas invaded by alien plants in the different provinces of South Africa (after Versveld et al., 1998). ............................................................................................... 34 Table 3.1: Results of the Principal Components Analysis. .............................................. 51 Table 3.2: Results of the Transformed Divergence (TDij) separability index.................. 54 Table 3.3 is a classification performance summary report of error matrices from the respective classifier algorithms. Note that the producer‘s and user‘s accuracy is specific to target class only. ...................................................................................... 60 Table 3.4: Commission and omission errors from the respective classifier algorithms generated from the image covering KZN midlands. Note that the errors of commission and omission are specific to A. mearnsii only. ..................................... 60 Table 3.5 is a classification performance summary report of BOX, MAHAL, ML and MDM. Note that the producer‘s and user‘s accuracy is specific to R. cunefolius only. .......................................................................................................................... 64 Table 3.6: Commission and omission error for ML, MDM, MAHAL and BOX algorithm. Note that the errors of commission and omission are specific to R. cunefolius only. ......................................................................................................... 64 Table 3.7is a classification performance summary report of BOX, MAHAL, ML and MDM. Note that the producer‘s and user‘s accuracy is specific to P. hysterophorus only. .......................................................................................................................... 66 Table 3.8: Commission and omission error for ML, MDM, MAHAL and BOX algorithm. Note that the errors of commission and omission are specific to P. hysterophorus only.................................................................................................... 68 Table 3.9 is a classification performance summary report of BOX, MAHAL, ML and DM. Note that the producer‘s and user‘s accuracy is specific to L. camara only. .. 69 Table 3.10: Commission and omission error for ML, MDM, MAHAL and BOX algorithm. Note that the producer‘s and user‘s accuracy is specific to L. camara only. .......................................................................................................................... 70 Table 4.1 is a lists and describing procedures executed in the successful mapping of selected IAP species .................................................................................................. 84 Table 4.2 Algorithm validation report............................................................................. 88 Table 5.1: Input parameters for ecological niche modelling. ........................................... 96 Table 5.2: Model assessment and validation results. ..................................................... 101
7
List of Abbreviations and Acronyms Al
Aluminium
AUC
Area Under the (ROC) Curve
CARA
Conservation of Agricultural Resources Act
CSIR
Council for Scientific and Industrial Research
DAEA
KwaZulu-Natal Department of Agriculture and Environmental Affairs
ESRI
Environmental Systems Research Institute Inc.
GIS
Geographic Information Systems
GPS
Global Positioning System
IAP
Invasive Alien Plant (species)
ISODATA
Iterative Self-Organising Data Analysis
KZN
KwaZulu-Natal
MD
Mahalanobis Distance
MAHAL
Mahalanobis (classifier)
MDM
Minimum distance to mean
ML
Maximum Likelihood
NASA
National Aeronautics and Space Administration
NDVI
Normalised Difference Vegetation Index
NEMBA
National Environmental Management Biodiversity Act
P
Phosphorus
RMSE
Root Mean Square Error
ROC
Receiver Operating Characteristic
RS
Remote Sensing
SAPIA
Southern African Plant Invader Atlas
SCOPE
Scientific Committee on Problems of the Environment
SPOT
Satellite Pour l‘Observation de la Terre
USGS
United States Geological Survey
WfW
Working for Water (Programme)
8
Executive Summary The problem of invasive alien plant species is of global significance, particularly in KwaZulu-Natal, the second most invaded province in South Africa. The varied biomes of KwaZulu-Natal make the province highly susceptible to invasion by a multitude of invasive alien plant (IAP) species, which not only threaten the biodiversity and water resource base of the region, but have dramatic impacts on human and animal welfare, livelihoods, and quality of life. Research is a key instrument to understanding and managing the invasion process, with effective and efficient strategic management protocols fundamental to curbing both current and future invasions.
Effective management and control of IAP species requires accurate and timely spatial information to delineate the location, spatial extent, and intensity of their invasion. Spatial technologies such as GIS (geographic information systems), GPS (global positioning systems), and RS (remote sensing), present a suite of tools and techniques that together afford a cost-effective, robust, integrated approach to IAP species management. This study explored the utility of these technologies for the management and control of invasive alien plant (IAP) species in KwaZulu-Natal, South Africa. The study aimed to use satellite remotely sensed data to map; develop algorithms for mapping and model the distribution of selected IAP species in KwaZulu-Natal.
Medium resolution SPOT 5 multispectral satellite imagery was tested in identifying and mapping four selected target species, namely, Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius at four selected study sites in KwaZulu-Natal; viz. midlands, coastal plain, northern plain, and highlands, respectively. A binary classification approach was adopted in training and classification of the satellite data. This included the use of unsupervised (ISODATA), and supervised (using four traditional classifiers viz. as Minimum-distance-to-mean, Parallelepiped, Mahalanobis and Maximum Likelihood algorithms) classification procedures.
The traditional classifier that yielded the highest classification accuracy was considered 9
for the development of the protocol and algorithm. The procedures used to map the chosen IAP species were developed into a protocol that was graphically modelled and executed as a semi-automated algorithm in ArcGIS-ArcEditor ModelBuilder.
The
algorithm was validated using imagery that differ both in terms of temporal and spatial coverage.
Habitat suitability of the four target species was further mapped across the province of KwaZulu-Natal using a presence-only data modelling approach. The model included eight physical and climatic data input layers that were recognised to be significant to plant growth and spread. All input data layers, coupled with species‘ presence data captured in the field, were modelled using the Mahalanobis distance statistic within a GIS environment.
The classification of SPOT 5 data to map P. hysterophorus produced overall accuracies of 0.54 (Kappa: 0.06), 0.50 (Kappa: 0.04) and 0.41 (Kappa 0.01) for the Maximum Likelihood,
Minimum-distance-to-mean
and
Mahalanobis
classfier
algorithms
respectively. Overall accuracies of 0.60 (Kappa: 0.19), 0.34 (Kappa: 0.00) and 0.47 (Kappa 0.03) were obtained using the Maximum Likelihood, Minimum-distance-to-mean and Mahalanobis classifier algorithms correspondingly in order to map R. cuneifolius. Mapping L. camara produced an overall accuracy of 0.74 (Kappa: 0.26) for the Maximum Likelihood classifier algorithm. In mapping A. mearnsii, 0.8 (Kappa: 0.6), 0.63 (Kappa: 0.25) and 0.63 (Kappa 0.25) overall accuracies were produced for the Maximum
Likelihood,
Minimum-distance-to-mean
and
Mahalanobis
classifier
algorithms respectively. The classification of imagery using the Parallelepiped classifier algorithm produced inferior results for the mapping of all selected species whilst the Minimum-distance-to-mean and Mahalanobis classifier algorithms produced similar results for the mapping of L. camara. The overall results conclude that the statistical classifier, Maximum Likelihood, is the ‗best‘ algorithm coupled with SPOT 5 data, for the mapping of the four IAP species targeted in this study. The distribution maps of the selected IAP species produced using the procedure Maximum Likelihood can be used for identifying the species with reasonable degree of accuracy 10
The algorithm and protocol developed for mapping the selected IAP species was designed such that, procedures necessary for the pre-processing of data such as atmospheric correction, geometric correction steps must be performed prior to using the algorithm. More importantly, generating training data of most if not all non-targets class within the scene is paramount.
The Mahalanobis distance statistic proved robust in predicting the habitat suitability (potential niche) of the four species modelled.
Overall accuracy ranged from 0.78
(Kappa: 0.67) for P. hysterophorus to 0.93 (Kappa: 0.75) for A. mearnsii. This was coupled with high specificity ranging from 0.65 (P. hysterophorus) to 0.81 (A. mearnsii). Calculation of Area under the Receiver Operating Characteristic (ROC) Curve (AUC) yielded high overall accuracy results ranging from 0.65 (L. camara) to 0.81 (A. mearnsii).
It is recommended that the utility of high spatial resolution multispectral data be evaluated in the identification and mapping of L. camara, P. hysterophorus, and R. cuneifolius. A higher spatial resolution coupled with a subpixel (fuzzy) classifier could improve the ability to discriminate between the target species and other vegetation, particularly grass. It is further recommended that the utility of hyperspectral data (e.g. EO-1 Hyperion) be explored for the identification and mapping of IAP species, particularly given the heterogeneity of the KwaZulu-Natal landscape.
The results from validating the developed mapping algorithm indicate that the mapping algorithm is best suited for semi-automatic mapping procedure, such that, the mapping algorithm requires training data that is specific to the image data being processed. The efficacy of using e-Cognition and syntactic approaches to classification vis-à-vis ArcGISArcEditor and statistics-based Maximum Likelihood classification to develop protocols and algorithms for mapping IAP species is recommended for investigation. Furthermore, a combination of Rule-based and syntactic approaches to classification offers the potential to develop classification protocols and algorithms, which may offer a possibility to off-set the need for generating training data. Without the need to generate training data, algorithms developed from such classification protocols may be fully automated. 11
The Mahalanobian model proved to be robust in delineating plant species‘ habitat suitability based on climatic and physical environmental attributes. It is however noted that a mapping resolution of 1.7km is ineffective at predicting local species‘ niches, and it is therefore recommended that finer scale data be acquired for fine scale mapping. Also, several new model approaches (e.g. neural networks, and boosted regression trees) are available that could be further investigated.
12
Chapter 1: General Introduciton 1.1
Introduction
‗Alien species are non-native or exotic organisms that occur outside their natural adapted ranges and dispersal potential‘ (Raghubanshi et al., 2005: 539). The World Conservation Union (IUCN, 2000) defines alien invasive species as organisms that become established in native ecosystems or habitats, proliferate, alter, and threaten native biodiversity.
Invasion by alien species is a global phenomenon, with threatening negative impacts to the indigenous biological diversity and ecosystems as well as related negative impacts on human health and well-being (Joshi et al., 2004; Richardson and van Wilgen, 2004), and the economy (Pimental et al., 2005). Invasion by plant species describe the uncontrolled proliferation and spread of a species to a new range, where it exhibits significant negative effects (NAS, 2002). Research is thus a key instrument in understanding and managing alien species and invasiveness (Pino et al., 2005), motivated not only by the need to mitigate the negative impacts of invasive species, but also to gain an understanding of the process of invasion underpinning fundamentals in ecological theory (Shea and Chesson, 2002).
Research on invasive alien plant (IAP) species in South Africa began as early as the 1930‘s (Table 1.1). During subsequent decades, research into management and control protocols for biological invasions was progressing (Richardson and van Wilgen, 2004), and has since become clear that the introduction of many woody IAP species into South Africa has lead to a multitude of negative, direct, and concomitant impacts. This is supported by the position that the South African government has taken in as far as IAP species are concerned; now requiring that all organs of state prepare monitoring, control, and eradication plans for areas within their jurisdiction (NEMBA, 2004). Consequently, at both national and regional levels, public and public-private initiatives to manage IAP species have been established. Regulations promulgated in terms of the Conservation of Agricultural Resources Act, Act 43 of 1983 (CARA) and the National Environmental Management Biodiversity Act, Act 10 of 2004 (NEMBA), and initiatives such as the 13
national ‗Working for Water Programme‘ (WfW) have been structured, funded, and tasked with the responsibility of prevention, eradication, containment, and control of all listed IAP species in South Africa (DEAT, 2009), with the impetus gained through social development, i.e. poverty alleviation, capacity building, and job creation (Cullis et al., 2007; Working for Water, 2004). In KwaZulu-Natal (KZN), the WfW is managed under the auspices of the Department of Agriculture and Environmental Affairs‘ (DAEA) Invasive Alien Species Programme (IASP).
Table 1.1: Key research initiatives on alien plant invasions in South Africa (adapted from Richardson and van Wilgen, 2004). Research programmes
Organization(s)
Duration
Examples of important scientific outputs1
Biological control of invasive alien plants
Department of Agriculture; Plant Protection Research Institute, Agricultural Research Council; Universities of Cape Town and Rhodes
Initiated 1930 – Ongoing.
Synthesis volumes
Catchment conservation research programme
South African Forestry Research Institute
1973 – 1990
Detailed studies on key invaders and invasion processes
South African National Programme for Ecosystem Research
Council for Scientific and Industrial Research (CSIR)
1977 – 1985
Regional syntheses
Scientific Committee on Problems of the Environment (SCOPE), programme on biological invasions
CSIR and many other organisations
1982 1986
Synthesis volumes
South African Plant Invaders Atlas
Plant Protection Research Institute, Agricultural Research Council
1975 – ongoing
Handbook; and detailed distribution studies
Working for Water programme
Department of Water Affairs and Forestry
1996 – Ongoing
First countrywide assessment of extent of woody plant invasions; Best-management practices proceedings
1
For details see Richardson and van Wilgen (2004).
14
In line with the advocacy of NEMBA, Act 10 of 2004 (Chapter 5), IAP species‘ programmes in South Africa have adopted a strategy which is to prevent, eradicate, contain, and control listed invasive alien plants. Each approach adopted is relevant to specific stages of invasion extent. As a consequence, many studies have focused on measuring the extent of invasion in the quest to better understand biological invasion in South Africa. However, each one of these studies focused on either a particular species and/or site in detail (Richardson and van Wilgen, 2004; van Wilgen et al., 2001). Clearly there still exists a need for multi-species nationwide surveys.
Biological control programmes for IAP species in South Africa were incepted in 1910 (van Wilgen et al., 2004), with current efforts influenced by the Convention on Biological Diversity, the South African government‘s integrated IAP species management efforts, and most significantly, the national Working for Water programme. Several studies have since been initiated; for example the control of Acacia melanoxylon by the seed-weevil (Melanterius acacia), and the use of a weevil species, (Melanterius maculates), for the control of Acacia mearnsii (Moran et al., 2000); and an investigation by Impson et al. (2004) on the use of a seed-feeding weevil (Melanterius servulus) for the control of Acacia cyclops.
Although biological control programmes may supplement current approaches to management and control strategies, there is much knowledge lacking with respect to weed ecology, geographic distributions, naturalisation and spread, phenology, invasiveness, and invasion history of IAP species in South Africa.
The need for
predictive, large-scale, fine-resolution mapping is further necessitated by hybridisation and evolutionary processes (Mooney and Cleland, 2001); the effectiveness of biological agents is reduced due to evolutionary pathways of resistance (NAS, 2002). This study fills, in part, the gap with respect to the geographic distribution of selected IAP species in KwaZulu-Natal and further aims to predict future invasions of already naturalised and current invader IAP species.
15
1.2
IAP Species: Challenges and Opportunities
The Southern African Plant Invader Atlas (SAPIA) (Henderson, 1998) is the most extensive nation-wide survey, having mapped IAP species at the scale of quarter-degree (15‘ scale) grids (Richardson and van Wilgen, 2004; van Wilgen et al., 2001). The database maintains approximately 50 000 records of more than 500 species, within the borders of South Africa, Lesotho, and Swaziland, dating back to 1979 (Henderson, 2001). This is, however, insignificant, as the data from the SAPIA database cannot be readily converted to spatial estimates of invaded areas (van Wilgen et al., 2001). Clearly, lack of nation-wide fine scale surveys (finer than SAPIA quarter degree grids) has implications to understanding species invasion and adopting management strategies. This necessitates the planning and implementation of well-timed, frequent, extensive, and finer-scale surveys. One aim of this study, therefore, was to attempt to map selected IAP species at a much finer scale (10m and 15m resolution) in KwaZulu-Natal. Results of this study may have a powerful influence on the future management of IAP species in KwaZulu-Natal, and in South Africa.
Broadly, studies on the impact of IAP species on ecosystem goods and services tend to be at a small spatial scale and confined to certain regions; the larger body of work having been confined to the fynbos biomes (Richardson and van Wilgen, 2004; van Wilgen et al., 2001). Consequently, lack of concrete nation-wide environmental assessments may lead to a poor estimate of the true economic impacts. The problem is exacerbated by the lack of an attempt to create an overall framework for the collection and synthesis of impacts at the scale of an ecosystem.
Containment and control of invasion by alien plant species constitute the core operations that most programmes working on alien plants are involved with in South Africa. In order to successfully contain and control alien plant species, strategies to address measuring and monitoring the distribution and densities of AIP species must be implemented. Evidently, frequent thematic maps are inherently associated with such undertakings.
16
1.3
Remote Sensing, Geographic Information System and IAP Species
Mapping Remote sensing (RS) has been used as a technique to measure and map vegetation (Aitkenhead and Aalders, 2000) and to map invasive alien plant species (Lawrence et al., 2006).
Given its many advantages, for example multi-temporal coverage and cost
effectiveness (Lillesand et al., 2008), and provision of a synoptic view of the earth‘s surface (Joshi et al., 2004), it is favoured as the tool of choice in natural resource and land use management, replacing traditional field surveys (Kokaly et al., 2003). The technology provides a practical approach to studying varying terrains, particularly inaccessible environments, provides a multitude of sensor systems at varied resolutions, alleviates spatial heterogeneity because of its broad view (Joshi et al., 2004), and attributed to the multi-date nature of digital imagery, is ideal for time-series analysis applications (Underwood et al., 2003).
Several studies have successfully exploited remotely sensed data to map vegetation, and to identify and map invasive alien plant species, for example: (1) Kokaly et al. (2003) achieved an overall classification accuracy of 74.1 % (Kappa index of 0.62) using airborne visible/infrared imaging spectrometer (AVIRIS) data to study vegetation in Yellowstone National Park; (2) classification of
spotted knapweed (Centaurea
maculosa) and leafy spurge (Euphorbia esula) using hyperspectral imagery, was undertaken by Lawrence et al. (2006) in southwest Montana rangeland in the United States, achieving an overall classification accuracy of 84% and 86%, with class accuracies of 60% and 93% for C. maculosa and E esula respectively; and (3) Narumalani et al. (2009) used AISA (Airborne Imaging Spectroradiometer for Applications) hyperspectral data to map four IAP species in North Platte River, Nebraska, achieving an overall map accuracy of 74%.
The above studies used
hyperspectral sensors and have high classification accuracies. However these are sensors not readily available in South Africa
The availability of SPOT 5 data coupled with the demanding need to effectively and efficiently manage, control, and eradicate IAP species in South Africa, while preventing 17
future invasions, justified the need for this study. Clearly, it is to be expected that resource managers would find themselves having to process large amounts of data prior to sanctioning action. More importantly, there is a clear necessity to act swiftly on the basis of information gathered when appraising IAP species. It is envisaged that the outcomes of this research will lessen the information technological gap in the invasive alien species programme, providing a robust approach to IAP species management.
A Geographic Information System (GIS) is a computer-based system for the capture, storage, retrieval, analysis, and display of spatial data. This spatial data is primarily acquired through RS and Global Positioning System (GPS), thus making remote sensing an integral component of a GIS (Lillesand et al., 2008; Skidmore, 2002).
The
combination of GIS and remote sensing techniques provides an opportunity to quickly map and model the distribution and spread of various vegetation types (Madden, 2004).
1.4
Motivation to the Study
This research is probably the first of its kind in South Africa; to identify, characterise, and map, the spatial distribution of selected invasive alien plants at a large spatial scale and at a fine resolution (10m and 15m), as well as model their habitat suitability in KwaZulu-Natal.
At the time of starting this study, the government departments and parastatals companies had access to June-July SPOT 5 imagery captured over a three year period. It is expected that the data volumes will gradually increase as more SPOT 5 imagery is captured and made available. In order to carry out timeous interventions in the fight against IAP species, these data sets will have to be processed at a relatively short period of time. Thus, part of the study is aimed at testing the utility of using SPOT 5 imagery to develop protocol and algorithms to map the selected major current and emergent IAP species. It is envisaged that the protocol and algorithm developed to map the selected major current and emergent IAP species, will help facilitate the processing and classification of SPOT 5. Research outputs (thematic maps and suitability models) are likely to serve a number of significant purposes as part of a spatial information systems (GIS) database. This 18
database will form an integral component of decision-making and support for organisations (e.g. the Invasive Alien Species Programme) involved in monitoring and control and/or clearing programmes as part of IAP species management. The key utilities of this database may be summarised as follows: i. Determining location and current spatial extent of IAP species ii. Monitoring invasive susceptibility, establishment, and infestation by new species, i.e. serve as early-warning system iii. Monitoring species distribution patterns and abundance iv. Risk evaluation v. Evaluation of efficacy of current, as well as future planning and implementation of monitoring and control programmes vi. Better-informed decision-making and implementation with regards to clearing programmes; support for prioritisation of activities, and areas for clearing vii. Establishing niche areas for further research: IAP species ecology, invasion and invasibility, risk management, monitoring and control.
1.5
Aim of Study
The overall aim of the project is to test the suitability of using utility of medium resolution multispectral SPOT 5 data to identify, map and monitor selected current and emergent invasive alien plant species in KwaZulu-Natal. The study also aims to develop protocols and algorithms thereof for identifying the selected current and emergent invasive alien plant species as well as explore spatial models to predict the distribution of potential future invasions using a geographic information system. The study further aims to explore spatial models to predict the distribution of potential future invasions using a geographic information system.
1.6
Objectives of Study
The specific objectives of the study were:
19
1. To identify and map selected major current and the emergent invasive alien plant species in KwaZulu-Natal, using remotely sensed SPOT 5 data. 2. To arrive at a protocol or framework for identifying and mapping invasive alien plant species and algorithms for identification. 3. To evaluate the spatial dynamics of IAP species invasions in KwaZulu-Natal using climate envelope modelling.
1.7
Description of Study Area
The study area for the investigation is the province of KwaZulu-Natal, between latitudes 26o and 31o S, and longitudes 28o and 32o E, located on the east coast of South Africa (Figure 1.1), shares borders with Lesotho, Swaziland and Mozambique (Fairbanks and Benn, 2000), and is approximately 88000km2 in spatial extent (Cooper and Cooper, 2002).
Geological formations run roughly north to south across an eastern-sloping
terrain, with eleven major fluvial systems draining eastward, cutting through the geological layers and resulting in a dissected topography, and deeply incised valleys (Cooper and Cooper, 2002).
Contrasting features of the KwaZulu-Natal province include altitudinal ranges from sea level to over 3000m in the peaks of the uKhahlamba Drakensberg, considerable temperature ranges; average winter temperature (13.9°C) being approximately 9oC cooler than average summer temperature (21.7°C), rainfall variability from 500-2000mm per annum, and a highly varied topography; undulating coastal plains to the rugged terrain of the midlands and sheer slopes of the uKhahlamba Drakensberg (Camp, 1997; Cooper and Cooper, 2002). The Indian Ocean, particularly the Agulhas current, greatly influences the climatic variability of the province; a subtropical coastal region characterised by high humidity, high temperatures, and high summer rainfall (900-1200mm); rainfall is mainly in summer (November to March), with resultant warm, wet summers and cool, dry winters (Fairbanks and Benn, 2000).
The great variation in climate, topography, and soils affect the range and distribution 20
patterns of vegetation in the province. The province maintains four biomes and twenty vegetation types; vegetation type being delimited by shared common species, similar vegetation structure, and shared ecological processes (Low and Rebelo, 1998). The lowlying hot and dry areas of northern KwaZulu-Natal, and most of the river systems, are dominated by bushveld, while various forest species are found across the high rainfall areas of the coastal belt, midlands mistbelt, highland sourveld and uKhahlamba Drakensberg.
Tall grassveld is characteristic of the northern plains with the cold
highland areas maintaining short grassland (Camp, 1997; Cooper and Cooper, 2002).
Figure 1.1: The study area of KwaZulu-Natal. Map (A) shows the four biomes of KwaZulu-Natal and the major fluvial systems draining the province, and map (B) indicates the four sites selected for the study lying within 4 vegetation types of KZN (after Low and Rebelo, 1998).
21
1.8
IAP Species Selection
Given several limitations of attempts to prioritise species, Nel et al. (2004) proposed classifying species into two categories, namely, ‗major invaders‘; species having been established in their environments and effecting substantial impact on ecosystem functioning, and ‗emerging invaders‘; species having lesser impact than the major invaders, however, maintaining the propensity of a major invader.
The approach adopted in this study was (1) to define a species list comprising both major and emerging invaders, (2) to prioritise these species according to their impacts and associated factors through the development of a species-prioritisation matrix (Appendix A), and (3) to arrive at a final list of four species through the use of a rigorous Multicriteria Evaluation method supported by an Analytical Hierarchy Process. The responses from experts were unfortunately less than expected, and the attempt was consequently abandoned. The information obtained was instead used to guide the preliminary field visits, which were significant in identifying and assessing the status of IAP species, and identify the dominant invader species in each of the selected study sites (Figure 1.2).
The final list included four species selected for the study; three major invaders (Acacia mearnsii, Lantana camara, and Rubus cuneifolius), and one emerging invader (Parthenium hysterophorus) species, at four study sites (midlands, coastal plain, northern plain, and highlands) in KwaZulu-Natal.
22
Figure 1.2: The four IAP species selected for the study. Clockwise, from top left: Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius.
23
1.8.1 Acacia mearnsii (current invader): KwaZulu-Natal midlands Black wattle (Acacia mearnsii De Wild.), a member of the family Fabaceae, is an evergreen leguminous tree, 6 – 20m tall, originating from South East Australia and Tasmania. A. mearnsii is documented as a transformer species and a declared invader (category 2) as legislated by CARA (1983). The species is a serious invader of veld, forest gaps, indigenous bush, roadsides, riparian zones, and occasionally, perennial crops e.g. sugarcane (Bromilow, 2001; de Neergaard et al., 2005; Henderson, 2001; Tassin et al., 2009). This study site lies within the ‗moist midlands mistbelt‘ of the grassland biome (‗short mistbelt grassland‘) of which approximately 89% has been transformed due to intensive agriculture and commercial forestry. The moist midlands mistbelt has an altitudinal range of 900-1400m, an annual rainfall range of 800-1280mm, and a mean annual temperature of 17.0°C (temperature varies between -2°C and 38°C).
Soils are
characterised by yellow/red apedal subsoils and organic-rich topsoils that maintain a high potential, but are of low nutrient status; being plagued by P-fixation and Al-toxicity, and are highly leached.
Few areas of indigenous vegetation remain; these areas being
populated by the dominant species Themeda triandra (Camp, 1997; Granger and Bredenkamp, 1998).
Field surveys documented this study site as being highly infested by Acacia mearnsii, which dominates the landscape as commercial forest plantations (predominantly SAPPI and Mondi2).
Significant to this study are the many occurring ―wild‖ A. mearnsii
clusters, which according to Govender (2007: pers comm.3), may be classified as a forest compartment if properly managed and a permit issued. These wild clusters encroach on rivers and streams, subsequently impacting on the water resources in the region. 1.8.2 Lantana camara (current invader): Coastal KwaZulu-Natal Lantana (Lantana camara L.), a member of the family Verbenaceae, is a compact 2
3
Plantation forest companies operating in KwaZulu-Natal. Govender, A., personal communication. Clearing team‘s area manager for Greytown, Umgeni Water.
24
floriferous shrub, growing up to 2m, native to tropical America. L. camara is considered one of the world‘s ten worst weeds and is one of the most serious invaders in South Africa and the Asian subcontinent. In South Africa, Lantana maintains the status of transformer and a classified category 1 plant (CARA, 1983). More than fifty variants have been documented; cultivated for ornamentals and hedging. Lantana invades forest and plantation margins, savanna, riparian zones, roadsides, and degraded land (Bromilow, 2001; Henderson, 2001; Sharma et al., 2005). This study site lies within the savanna biome (‗coastal bushveld-grassland‘). Much of this site has been transformed due to agriculture and forestry. This study site lies between sea level and an altitude of approximately 450m, and has a rolling to flat terrain. Rainfall is primarily in summer, with a mean annual range of 820-1423mm. Summers are typically hot, coupled with mild winters; mean annual temperature range of 22.0°C and high humidity. Soils are characteristically sandy, of Quaternary Aeolian and marine origin, and not of high agricultural potential. Sugar cane and exotic tree plantations dominate the landscape; most natural vegetation having been cleared. Some remnant coastal forest species (e.g. Millettia grandis, Strelitzia nicolai, and Syzygium cordatum) remain while grasslands are scrubby, maintaining some tree species including Acacia karroo, Acacia nilotica, and Acacia robusta (Camp, 1997, Granger et al., 1998).
Field surveys revealed high infestations of Lantana camara in this study site. L. camara is quickly establishing itself, although competing with Chromoleana odorata (Sihlangu, 2007: pers comm.4). L. camara occurs in dense impenetrable thickets, particularly in areas that have been recognised as abandoned Eucalyptus spp plantations.
1.8.3 Parthenium hysterophorus (emergent invader): Northern KwaZulu-Natal Parthenium weed (Parthenium hysterophorus L.), a member of the family Asteraceae, is an erect ephemeral herb, native to sub-tropical North and South America, and the Caribbean. P. hysterophorus degrades natural ecosystems, invading agricultural and rangeland habitats, outcompeting native species through allelopathy. Parthenium weed 4
Sihlangu, S., 2007. Personal Communication. Clearing team‘s area manager, Richards Bay-Empangeni, Umhlathuze Municipality.
25
maintains its legal status as a category 1 invader (CARA, 1983). P. hysterophorus invades roadsides, watercourses, cultivated fields, and overgrazed lands, and affects both humans and animals, being a skin and respiratory tract irritant. This study site lies within the savanna biome (‗Lebombo arid mountain bushveld‘), lying below 450m and topographically flat to undulating, except for the slopes of the Ubombo Mountains (27.382°S; 32.015°E). This site experiences a mean annual rainfall of 450700mm and a mean annual temperature of 23°C (temperature varies between -1°C and 46°C). The geology of the region is rhyloite and granophyre, giving rise to lithosols; shallow, acidic, sandy soils. The characteristic vegetation is an Acacia nigrescensSclerocarya-Themeda Savanna (Camp, 1997; van Rooyen and Bredenkamp, 1998).
Field surveys revealed high infestations of P. hysterophorus, predominantly along roadsides, in cattle dips, schoolyards, and in poorly managed vegetable gardens.
1.8.4 Rubus cuneifolius (current invader): KwaZulu-Natal highlands American bramble (Rubus cuneifolius Pursh), a member of the family Rosaceae, is a sprawling, thorny shrub of North American origin.
Several species of bramble are
indigenous to South Africa, e.g. R. rigidus, while R. cuneifolius is a declared weed (CARA, 1983) and recognised transformer. R. cuneifolius is becoming a serious invader of grasslands, forest edges, plantations, roadsides, and riparian zones, particularly in parts of KwaZulu-Natal where it forms impenetrable clumps (Bromilow, 2001; Henderson, 2001). This study site lies within the grassland biome (‗moist upland grassland‘) characterised by moist, cold montane grasslands, at an altitude of 600-1400m. Rainfall is mainly limited to the summer months; a mean annual rainfall of 650-1000mm and a mean annual temperature of 16°C (temperature varies between -3°C and 40°C). Soils are relatively deep, rocky, highly leached, and acidic, derived from Karoo Sequence sediments and dolerite. The landscape is dominated by fire-maintained grassland; species include, but not limited to, Alloterpsis semialata, Themeda triandra, Sporobolus africanus, as well as 26
several species of Eragrostis (Bredenkamp et al., 1998; Camp, 1997).
Field surveys revealed high infestations by Rubus cuneifolius throughout the study site, particularly along roadsides and on poorly managed gardens.
In Kamberg Nature
Reserve, central uKhahlamba Drakensberg (29.40°S; 29.67°E), R. cuneifolius occurs within the reserve and neighbouring privately owned farms as relatively large, homogenous, impenetrable clumps. R. cuneifolius also appears to be an opportunist, fast encroaching on areas where A. mearnsii has been cleared.
1.9
SPOT 5 Satellite Imagery
Satellite Pour l‘Observation de la Terre (SPOT) is a system of satellites in operation since 1986, with the launch of SPOT 1. SPOT 5, launched in May 2002, follows a polar, circular, sun-synchronous orbit at an altitude of 822km, and maintains an orbital cycle of 26 days.
SPOT 5 comprises two high resolution geometric instruments with a spectral range from the visible (0.50-0.68µm) to the shortwave infrared (1.58-1.75µm). Two panchromatic bands (0.48-0.71µm), with a spatial resolution of 5m, are combined to generate a 2.5m product. Each scene has a swath of 60km, and a spatial resolution of 10m (multispectral), and 20m (SWIR). The instrument has an absolute location accuracy of 30m (no ground control points; flat terrain).
1.10 Organisation of the Report Chapter 2 provides an overview of the problem of invasive alien plants and the utility of the spatial technologies as an integral part of management and control programmes and protocols. The use of remote sensing and GIS is reviewed for IAP species mapping, automation of mapping procedures as well as modeling.
Chapter 3, 4 and 5 addresses the three many objectives of the study, and have been 27
reported as stand-alone studies.
Chapter 3 evaluates the utility of 10m multispectral SPOT 5 to generate thematic maps of Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius infestations.
Two primary approaches adopted in training and classification of the
datasets are designed: unsupervised (ISODATA) and supervised (using four traditional classifiers viz. as Minimum-distance-to-mean, Parallelepiped, Mahalanobis and Maximum Likelihood algorithms) procedures.
Chapter 4 explores the utility of automating the mapping of Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius using pixel-based classifier algorithms
Chapter 5 investigates an explicit spatial modelling approach to predicting invasion susceptibility, utilising the Mahalanobis distance statistic to generate habitat suitability models within a GIS environment. The Mahalanobian model is of particular significance given its robustness over rectilinear envelope models.
Finally, Chapter 6 provides a general discussion, conclusions, and recommendations of the report.
28
Chapter 2: Literature Review 2.1
Introduction
Invasive alien plant species (commonly referred to as weeds) have emerged as a persistent problem in South Africa.
They have dramatic impacts on the natural
environment and concomitant implications for human welfare, livelihoods, and quality of life.
Biological invasion is a natural process; however, human intervention has
accelerated the rate of spread and naturalisation of many species across a multitude of foreign landscapes (Ewel et al., 1999), which has significantly increased during the last two centuries. Several studies have recognised humans and/or human activities as the driving force behind the introduction, and spread of alien species across ecosystems, with a direct correlation between the intensity of human activities and the intensity of IAP species invasion (Frenot et al., 2001). Humans are consequently both purposely and inadvertently vectors for the spread of IAP species across South Africa, and across the planet. ‗Disturbance is a natural and oftentimes integral occurrence in most ecosystems‘ (Pritekel et al., 2006:1). Invasibility by an invasive alien plant species represents a unique form of disturbance (Pritekel et al., 2006), which may be attributed to factors such as the invaded area‘s climate, the environment‘s disturbance regime, and the competitiveness of the native species; the actual invasion being influenced by the number of invading species, their biology, and the propensity of the invaded environment to be invaded (Lonsdale, 1999).
The spread of invasive alien plant species presents a real threat to global biodiversity and ecosystem functioning (Mooney and Cleland, 2001) and is probably second only to that of outright habitat destruction (Ragubanshi et al., 2005). The threat is likely to escalate given continued human-induced disturbances, for example habitat fragmentation resulting from agriculture and global climate change (Ewel et al., 1999), with humans not taking cognisance of, and finding mitigatory measures to, their negative consequences. Given their eminent threat to biodiversity and ecosystem functioning, alien invasive plant species are of great concern and of focal interest to natural scientists and natural resources 29
managers alike. Many hypotheses have tried to explain the variability in invasiveness between different plant communities, however there is currently no general theory to explain community invasibility (Lonsdale, 1999).
Davis et al. (2000) theorise that
invasibility is a result of fluctuations in resource availability, while Thompson et al., (2001) recognise that invasibility is a product of an increase in availability of resources, either through increase or reduction of resources. Studies by Thiébaut (2005) on Elodea species, suggest that invasion corresponds primarily to increased resource availability.
Invasiveness of a species involves complex interactions with the invaded environment (Kolar and Lodge, 2001); invasive species often modifying the invaded environment thereby making it more hospitable (Cuddington and Hastings, 2004).
Several
characteristics of invading species, for example, allelopathy, fire tolerance, competitive ability, vegetative reproduction, and fitness homeostasis and phenotypic plasticity (Sharma et al., 2005) have been identified as predictors of invasibility. Invasion success of an IAP species is dependent on several factors including species‘ characteristics, characteristics of the invaded environment, and species‘ composition of the invaded environment (Kotiluoto, 2008).
2.2
IAP Species in Context: Impact on the Natural Resource Base
‗Commercial forestry based on alien trees is a well established feature of the South African landscape and economy‘ (Le Maitre et al., 2002:144), with species of pines and eucalypts, having been introduced for timber production (Rouget et al., 2004a), covering approximately 1.5 million ha. Benefits of these alien plantations include economic development, employment, and foreign exchange through export of forestry products (Le Maitre et al., 2002).
Unfortunately, these plantations have been associated with
considerable negative impacts, most notably, significant reductions in streamflow, and substantial impacts on biodiversity and functioning of ecosystems (Le Maitre et al., 2002). Of all the alien forest plantations, black wattle (Acacia mearnsii), silver wattle (A. dealbata), blackwood (A. melanoxylon), bluegum (Eucalyptus globulus), and cluster pine (Pinus pinaster) have the greatest impact on water resources (van Wilgen et al., 1997). 30
Le Maitre et al. (2000) report a strong correlation between commercial forest stands and alien invasive plant species invasion in South Africa, with Nel et al. (2004) reporting a 44% overlap between IAP species and commercial forestry. Seventy eight percent of the 2.9 million ha invaded by Pinus species are attributed to the forestry sector; Eucalyptus accounts for 2.4 million ha of invasion, of which 37% attributes to forestry; and A. mearnsii has invaded an area greater than 2.4 million ha, of which 10% accounts for forestry. South Africa expands some 1 221 040 km2, all of which is particularly prone to invasion by alien plant species (Table 2.1) (Versfeld et al., 1998). Latest reports suggest that about 10 million ha, or 8.28% of South Africa has already been invaded (Le Maitre et al., 2000; Moran et al., 2000) by more than 180 invading plant species, consuming 3.3 billion litres of water in excess of native species, accruing to a loss of approximately 6.7% of mean annual surface runoff (Le Maitre et al., 2002; Moran et al., 2000). MacDonald et al. (2003) asserts that ~750 000 ha of invaded land needs be cleared annually if the battle against invasive plants is to be won by 2023. This 20-year effort would, however, come at a projected cost of R5.5 billion (Le Maitre et al., 2002).
Table 2.1: Top 10 invading species in South Africa ranked by condensed invaded area (after Versveld et al., 1998). Species Acacia cyclops Prosopis spp. Acacia mearnsii Acacia saligna Solanum mauritianum Pinus spp. Opuntia spp. Melia azedarach Lantana camara Hakea spp.
Habitat landscape alluvial plains riparian/landscape riparian/landscape riparian/landscape landscape landscape riparian/landscape riparian landscape
Total invaded area (ha)
Condensed invaded area (ha)
1 855 792 1 809 229 2 477 278 1 852 155 1 760 978 2 953 529 1 816 714 3 039 002 2 235 395 723 449
339 153 173 149 131 341 108 004 89 374 76 994 75 356 72 625 69 211 64 089
Density (%) 18.28 9.57 5.30 5.83 5.08 2.61 4.15 2.39 3.10 8.86
Studies based on the national Working for Water Programme (WfW) indicate that the mean cost of initial clearing and follow-up of a 75–100% IAP density class is 31
approximately R1 000 per ha, excluding the cost of herbicides (Marais et al., 2004). This equates to hundreds of millions of Rand being allocated annually for IAP management and clearing programmes. Expenditure for WfW operations increased from R25 million during 1995/96 to more than R400 million during 2003/04 (Marais et al., 2004). A recent evaluation of costs of clearing (excluding biological control) has been estimate to be approximately R1.6 billion (Cullis et al., 2007).
IAP species present a serious problem in fynbos in the Western Cape (Figure 2.1), with species of Acacia, Eucalyptus, Hakea, and Pinus having altered the biophysical nature of this biome (van Wilgen, 2009).
Van Wilgen (2009) further highlights that fynbos
invaded by pines and wattles are frequently subject to fires.
Altered fire regimes
(unplanned wildfires) are highly problematic, as fires affect the rate of spread of some species, by triggering the release and subsequent germination of seeds resulting in an increase in density and extent of infestation (Brooks et al., 2004; van Wilgen, 2009).
KwaZulu-Natal is the wettest province having a mean annual rainfall of 553–1353 mm and a mean annual runoff of 20–678 mm. The varied biomes of the province (Figure 2.1) increase its susceptibility to invasion by a multitude of IAP species (Table 2.2), which severely impact on the provinces‘, and country‘s water resources (Versfeld et al., 1998). Invading alien plant species have been recorded to be consuming an extra 576 million m3 of water per annum compared to previously indigenous vegetation, with Umgeni Water (1997) having reported an increased 11 million m3 per annum consumption by invading species in riparian zones above the Midmar Dam. This situation has probably worsened over the past decade, although no literature surveyed has documented this.
32
Figure 2.1: The varied biomes of South Africa (map A) and KwaZulu-Natal (map B).
33
Table 2.2: Most important invader species in KwaZulu-Natal ranked by condensed invaded area (after Versveld et al., 1998). The condensed area is the total area adjusted to bring the cover to the equivalent of 100%. Species Solanum mauritianum Chromolaena odorata Acacia (mixed species) Acacia dealbata Acacia mearnsii Rubís spp Lantana camara Pinus spp Eucalyptus spp
Habitat
Total Condensed invaded area invaded area (ha) (ha)
riparian riparian riparian/landscape riparian/landscape riparian/landscape riparian riparian riparian/landscape riparian
404 999 326 139 116 584 136 004 190 542 163 475 235 849 49 810 153 103
44 265 43 178 24 912 22 115 12 896 11 845 10 518 4 115 2 658
Density (%) 10.93 13.24 21.37 16.26 6.77 7.25 4.46 8.26 1.74
KwaZulu-Natal is notably the fifth most invaded province by total invaded area (Table 2.3), the Western Cape being the most invaded, and Gauteng, with the least total invaded area (Le Maitre et al., 2000). In KwaZulu-Natal (9 459 590 ha), alien plants (most notable A. mearnsii and A. dealbata) have invaded an area of 922 012 ha, equivalent to 9.75% of the province, with the coastal zone and midlands recognised as major problem areas (Le Maitre et al., 2000). Solanum mauritianum is the most condensed invader while the acacias are most widespread (Versfeld et al., 1998). Table 2.3: Areas invaded by alien plants in the different provinces of South Africa (after Versveld et al., 1998). Province Eastern Cape Free State Gauteng KwaZulu-Natal Mpumalanga Northern Cape Northern Province North West Western Cape South Africa
Area (ha) 16 739 817 12 993 575 1 651 903 9 459 590 7 957 056 36 198 060 12 214 307 11 601 008 12 931 413 121 746 729
Total invaded area (ha) 671 958 166 129 22 254 922 012 1 277 814 1 178 373 1 702 816 405 160 3 727 392 10 073 908
(%) 4.01 1.28 1.35 9.75 16.06 3.26 13.94 3.49 28.82 7.99
Condensed invaded area (ha) % 151 258 24 190 13 031 250 862 185 149 166 097 263 017 56 232 626 100 1 735 856
0.90 0.19 0.79 2.65 2.33 0.46 2.15 0.48 4.84 1.37
34
The grassland biome of KwaZulu-Natal is heavily invaded by several tree species, including species of Acacia, as well as woody species such as Chromolaena odorata and species of Rubus; with riparian zones being the most severely invaded (Richardson and van Wilgen, 2004). A report by Umgeni Water (1997) indicated that the abundance of Acacia species may in fact be considerably higher, were more detailed mapping become available. Umgeni Water (1997) further point out that 40% of riparian zones (2056km mapped) within three quaternary catchments above Midmar Dam, were invaded by species of wattle.
Alien invasion does not follow a set pattern (Sanz-Elorza et al., 2006), is not proportionate in its global expansion (Sanz-Elorza et al., 2006), nor is the distribution of alien plants equal in any given area (Thiébaut, 2005). The continued propagation and establishment of IAP species across new ranges is causing homogenisation of flora (Mooney and Hobbs, 2000), loss of biodiversity (Moran et al., 2000; NAS, 2002), changes in disturbance regime (Brooks et al., 2004), reduced runoff and streamflow (Moran et al., 2000; NAS, 2002), changes in nutrient cycling (NAS, 2002), affecting soil surfaces by altering germination sites and surface micro-climates (Pritekel, et al., 2006), causing severe erosion and degradation of soils (Brooks et al., 2004), and is considered to be a significant driving force behind global change and species extinction (Richardson and van Wilgen 2004).
2.3
IAP Species in Context: Impact on Human Welfare and Quality of Life
An interesting definition of a weed is provided by Bridges (1994:392): ‗A weed is a plant growing where man (a person) wishes other plants, or no plants, to grow and which has some economic, ecological, or aesthetic implication for man (a person) and/or his (or her) activities‘. Globalisation has opened avenues for international trade and transport, creating corridors for emigration of species into non-native environments (Ewel et al., 1999). Explosion in human population has seen increased movement of people around the world with concomitant increase in the spread of IAP species (Frenot et al., 2001; Lonsdale, 1999; Thuiller et al., 2006). 35
Humans are bound to continue to further introduce non-indigenous species into local plant populations, for example, alien species may form the basis for maintaining productivity in agricultural ecosystems, and for horticulture (Ewel, 1999), while other alien species, for example, Casuarina equisetifolia (beefwood) is utilised by the mining industry for dune stabilisation, though it is a serious invader of coastal dunes and sandy sea-shores (Henderson, 2001). Humans continue to exploit IAP species; timber from pines, tannins from species of acacias, and aesthetic and ornamental value of Jacaranda mimosifolia and Lantana camara (Chapman and Le Maitre, 2001).
To the local
communities in the rural parts of the Drakensberg region of South Africa, two IAP species, namely, A. mearnsii and A. dealbata, represent a natural resource—a primary heat source, building material, medicinal extract, and a source of income from the sale of firewood (de Neergaard et al., 2005). The duality of the situation is, however, the burgeoning need to clear ‗wild‘ wattle stands.
As noted earlier, forest plantations of alien species have been around since the late 1800s. Chamberlain et al. (2005) reported that commercial forests covered approximately 1.37m ha of South Africa‘s landscape, of which 543,210ha (~40%) is grown in KwaZulu-Natal. In addition to posing detrimental effects to biodiversity through decreased water availability, serious consequences arise from greatly reduced surface runoff and streamflow (Enright, 2000).
In a water scarce country like South Africa, reduced
streamflow, and water levels in dams and reservoirs, threaten the socio-economic status of these regions; rivers and dams maintain an important source of water for human consumption, irrigation for agriculture, and open water for recreation (Le Maitre et al., 1996). The forestry, timber, pulp, and paper (FTPP) industries do however, contribute 1% (~R12.2bn) to the GDP, and employ approximately one hundred and seventy thousand people (Chamberlain et al., 2005).
Related impacts of IAP species on humans and human activities include decreased agricultural productivity with concomitant impacts on both crops and animals. This form of impact relates directly to economic impacts accrued through both loss and cost, and may be in the region of billions of Rand (Bridges, 1994). There is also documented 36
evidence of adverse effects to both human and animal health and well-being, for example, Parthenium hysterophorus is known to cause asthma and dermatitis in humans (Bromilow, 2001) and ingestion by animals (livestock and wildlife) may have far reaching consequences with respect to reproduction, production, and general wellness (Bridges, 1994). Bridges (1994) further points out that IAP species generally create an aesthetically displeasing environment, which often leads to loss in recreational potential.
2.4 Detection and Mapping of IAP Species using Remote Sensing Rejmánek (2000) highlight three fundamental management objectives for IAP species, namely,
prevention/exclusion,
early
detection/rapid
assessment,
and
control/containment/eradication, which in theory, might be very simple and straight forward to accomplish. Meeting these objectives, however, given the socio-economic and political climate of South Africa, is more a question of policy and technology (Rejmánek, 2000).
To effectively manage and control invasive non-indigenous plant species, and protect and preserve the local biodiversity and ecosystem functioning, managers require accurate and timely spatial information to delineate the location, spatial extent, and intensity of the invasion (Lawrence et al., 2006; Narumalani et al., 2009). This spatial information, acquired through remote sensing (RS) techniques, assists managers in monitoring the efficacy of current management and control strategies, monitoring possible future invasions (risk assessment), and assists in identifying target species and areas for clearing (Underwood et al., 2003).
Remote sensing may be defined as the acquisition of data about an object(s) on the surface of the earth, without the sensor being in physical contact with the object(s) itself. Currently available RS systems include air-borne, satellite/space-borne, radar (radio detection and range), and lidar (light detection and ranging); the specification of the system employed being dependent on the intended use and data requirements of the investigation. 37
Image resolution is an important consideration in mapping vegetation, as spatial as well as spectral resolution impact on the accuracy of the mapping (Lawrence et al., 2006). Aerial photography can produce very high spatial resolution imagery (Lawrence et al., 2006), is relatively inexpensive, with large archival data available (Underwood et al., 2003), can readily provide spatial estimates of invasion, and quantify infestation and spread (Müllerova et al., 2005). Several studies have applied digital aerial photography to vegetation mapping, for example, Anderson et al. (1996) used colour and colour infrared photography to detect infestations of leafy spurge (Euphorbia esula) in Theodore Roosevelt National Park, Rowlinson et al. (1999) used black and white, and colour aerial photography to map riparian vegetation in KwaZulu-Natal (overall accuracy: 57.61% 68.74%), and Müllerova et al. (2005) used aerial photographs (panchromatic, multispectral, and orthophotographs) to monitor the spread dynamics of Heracleum mantegazzianum in the west Bohemia.
Unfortunately, aerial photography requires extensive manual digitising, geo-registering, and interpretation and is not feasible for data collection over a large spatial area (Underwood et al., 2003), may not be cost-effective where repeated coverage is required as part of a monitoring programme (Müllerova et al., 2005), classification may be highly subjective, as class discrimination is based on the analyst‘s opinion (Rowlinson et al., 1999), and more significant, aerial photography has limited application to mapping individual species as vegetation reflectance cannot be readily spectrally delineated (Okin et al., 2001). It is for these reasons that satellite remote sensing is increasingly exploited in environmental studies.
2.5
Automated Procedures for Image Classification: Protocols and
Algorithms for Mapping IAP Species The introduction of software that combines remotely sensed data and GIS coupled with image processing functionality not only provides users with in-built tools but, they also provide users the possibility of creating customized procedures for image processing that can be automated. In its most general sense, an algorithm is somewhat similar to protocol 38
since it also details instructions which results in a desired outcome (.Nixon and Aguado, 2000). However, algorithms in computers execute a set of procedures or instructions automatically.
Basically, procedures are automated or developed into algorithms
primarily because it may be difficult to keep track of the various datasets, processing procedures, parameters and assumptions used during analysis (ESRI, 2000).
The
introduction of commercial or proprietary software on the other hand (e.g. remote sensing and geographic information systems (GIS) commercial software packages) not only provides users with in-built tools but they also provide users the possibility of creating customized procedures for image processing.
More importantly the introduction of a user-friendly visual programming language (graphical modeling) has made it possible for non-scripting language experts to create customized and automated procedures for image processing. It is to this technological advancement coupled with the need for timely produced thematic maps that an investigation into development of an automated protocol for mapping target AIP species is conducted.
ERDAS and ArcGIS-ArcEditor which are commercial software packages were tested for development algorithms. These commercial packages were considered primarily since they provide users with both image processing toolkit and the possibility of the creating user defined procedures for applications. Secondly, the interface for creating customized procedures is not only operated using language syntax, but it also uses user-friendly visual programming language or flow-charts diagrams or symbols representing or modelling procedures involved such that they can be saved and used for automating procedures or creating algorithms (ESRI, 2000; Leica Geosystems, 2006; Gonzalez et al., 2004). Finally, and of most significance, the tested software for protocol and algorithm development was determined in part by access and availability on the part of the researchers and the stake holder.
39
2.6
Modelling IAP Species
Some of the key questions addressed by a GIS include location (what is at…?), condition (where is it…?), distribution (what is the distribution/pattern…?), trend (what has changed…?), and routing (which is the best way…?) (Skidmore, 2002). Given the analytical capabilities of a GIS, it is widely and increasingly utilised as a planning, management, and decision-support tool, particularly within the environmental and conservation sciences (Skidmore, 2002).
An increasing number of studies have already applied GIS to mapping and modelling the distribution and spread of IAP species (Joshi et al., 2004). For example, Rouget et al. (2004a) used GPS technology during field surveys for predictive mapping of Pinus species in the Kango Valley, in the Western Cape, South Africa, Mau-Crimmins and Orr (2005) successfully integrated GIS as part of a ‗GIS/GPS for Weeds Mapping‘ programme in Arizona, USA, and Boylen et al. (2006) used GIS to simulate the spread of Myriophyllum spicatum, Trapa natans, and Potamogeton crispus using herbarium records and documented dates of sightings, in New York State, USA. Gillham et al. (2004) developed an algorithm which was incorporated into a GIS; the Weed Invasion Susceptibility Prediction (WISP) model, for the prediction of locations of weed infestation.
A GIS enables the quantification of relationships between the variables (climatic, edaphic, and land management protocols) that determine invasibility, and the observed abundance of invader species, subsequently supporting the development of models of species distribution and spread (Bradley and Mustard, 2006; Vaclavik and Meentenmeyer, 2009).
A GIS is now a significant component of IAP species
management programmes; its effectiveness significantly improved with the integration of global positioning system (GPS) data (Travaini et al., 2007). A GPS5 provides accurate positional data which can be seamlessly integrated into a GIS for subsequent analysis and information production (Gillham et al., 2004; Nel, 2008). This system would thereby allow for more accurate surveys and mapping. 5
For a review of GPS see Nel, 2008.
40
GIS is now utilised extensively in the field of ecological surveying and management (e.g. Anderson and Martínez-Meyer, 2004; Draper et al., 2003), presenting a coherent set of tools for the input and analysis of spatial data (Chang, 2010) and for the modelling of complex habitats (Skov and Svenning, 2003).
Monitoring and mapping of the
biophysical and ecological characteristics of several invasive species has successfully been accomplished through the utility of integrated remote sensing and GIS technologies (Los et al., 2002; Rowlinson et al., 1999). Madden (2004) concludes that integrated remote sensing and GIS technologies (with the added facility of GPS data integration) have proved useful in planning and strategising of IAP species management (detection – mapping – control/eradication).
Joshi et al. (2004) emphasize the need for accurate and timely assessment and modelling of IAP species disturbance patterns if we are to successfully map the type and extent of plant invasions, as well as predict their potential impacts and/or risk to their new range. For several years, Geographical Information Systems (GIS) methodologies and techniques have been applied to understanding and mapping the relationships between the invasion, naturalisation, and spread of IAP species in new environments, and their spatial heterogeneity across ecosystems (Pino et al., 2005). Predictive modelling of species‘ distribution is a significant component of IAP species management programmes (Browning et al., 2005; Anderson et al., 2003). These models link presence (and absence) data to predictor variables (biotic and abiotic environmental factors) to predicting spatial patterns of invasion (Anderson et al., 2003; Bradley and Mustard, 2006; Lodge et al., 2006), through environmental matching (niche modelling) (Lodge et al., 2006). GIS-based spatial distribution models, coupled with advanced geospatial and statistical techniques, enable the prediction of IAP species location and distribution across landscapes (Vaclavik and Meentenmeyer, 2009).
GIS further
translates models to maps that can readily be used by resource managers for planning and implementation of monitoring and control programmes (Travaini et al., 2007).
41
Hellgren et al. (2007) suggested 5 basic steps to GIS modelling: i.
Extract descriptive habitat data (e.g. temperature, rainfall, and elevation)
ii.
Statistical analysis (quantifying relationships between presence/adsence data and predictor variables)
iii.
Spatial modelling (in GIS based on statistical analysis)
iv.
Mapping and simulations
v.
Model testing and validation (accuracy assessment, sensitivity analysis, uncertainty analysis, cross-validation)
These five steps essentially form the basis for habitat niche modelling in a GIS environment (e.g. Clark et al., 1993) and form the basis of the habitat suitability modelling in this study.
Traditional approaches to predicting IAP species distribution in a GIS environment include multivariate statistical methods including discriminant analysis, multiple regression, and logistic regression (Lodge et al., 2006). New approaches have since been developed, for example BIOCLIM (e.g. Beaumont et al., 2005), MD (Mahalanobis distance) (e.g. Calenge et al., 2008), CLIMEX (e.g. Sutherst et al., 1999), ENFA (Ecolgical Niche Factor Analysis) (e.g. Hengl et al., 2009), GARP (Genetic Algorithm for Rule-Set Production) (e.g. Anderson et al., 2003), and MAXENT (e.g. Phillips et al., 2006), and SPECIES (e.g. Pearson et al., 2002).
In a comparative evaluation of six of these modelling approaches, Tsoar et al. (2007) showed that GARP and MD proved superior over the other models, with a Kappa of 0.554±0.280 and 0.540±0.286 for the GARP and MD models respectively.
The
modelling approach adopted in this study is the MD; applied to predicting the spatial distribution of Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius in KwaZulu-Natal, South Africa.
42
2.7
Conclusions
The effective and efficient management of invading species certainly demands an integrated, strategic, well-structured, and coherent approach if we are to one day obviate such a threat. Spatial technologies afford a suite of tools that provide for an integrated approach to biodiversity, environmental appraisals, and land and resource management.
Albeit the extensive use of RS across many parts of the world, its use in the management of alien invasions in South Africa has been largely unexploited. This may be largely attributed to a general lack of knowledge and skills in the field, and predominantly the accessibility, availability, and high associated cost. The future does hold much promise though, with a multitude of new sensors being launched by many countries, and the costs associated with acquisition and processing of data being greatly reduced, and at times, available at no cost (for example archival Landsat, and ASTER for research purposes).
Coupled with the availability of multi-spatial (30cm to 1km) and multi-spectral (multispectral and hyperspectral) resolution data from a multitude of sensors, and increased computation power and efficiency, has been the advancement of image processing techniques. Latest techniques include classification and regression trees (e.g. Lawrence et al., 2006), artificial neural network models (e.g. Mas and Flores, 2008), subpixel (fuzzy) classifiers (e.g. ERDAS, 2008c), and object-oriented image analysis (e.g. Stow et al., 2009). These algorithms show great promise for the effective extraction of information from spectral data in a wide array of applications.
Since the South African strategy towards IAP species is to prevent, eradicate, contain, and control listed invasive alien plants (DEAT, 2009), part of the overall strategy should address measuring and monitoring the distribution (and rate of spread), and densities of IAP species across the country. Shortcomings of the SAPIA database and the Versfeld et al. (1998) study necessitate reappraisals and/or new investigations. Frequent large-scale thematic maps are inherently associated with such undertakings, especially for containment and control stages of invasion, which are the core operations that most IAP species programmes are involved with in South Africa. 43
One of the objectives of this study was to comment on the utility of remote sensing and GIS as integrated tools for the effective and efficient management and control of IAP species in South Africa. All the literature on IAP species reiterate the need for effective management strategies, while commenting on the utility of RS and GIS for the mapping and modelling of IAP species distribution and spread, across both time and space. This provided further impetus for this research into the utility of RS and GIS as strategic IAP species management tools, and its integration into management and control programmes in South Africa.
44
Chapter 3: Mapping Invasive Alien Plant Species 3.1
Introduction
Management and control programmes of IAP species cost billions of Rand annually (Pimentel et al., 2005); prevention being the best, and most cost effective approach (Kaiser, 1999), while agencies vested with this responsibility struggle with budget, information, and tools constraints (Barnett et al., 2007).
To date, a multitude of
information and mapping systems have been developed (Barnett et al., 2007) which include information on distribution, abundance, and habitat (Henderson, 1999). The Southern African Plant Invaders Atlas (SAPIA) (Henderson, 1999) is the most widelyused information system of its kind (species mapped at the scale of a quarter degree square), forming the basis of most studies in South Africa.
Given the limitations of the SAPIA database (van Wilgen et al., 2001) the opportunities presented by satellite remotely sensed data may be exploited to provide up-to-date information with regards to IAP species‘ locations and distributions across varied landscapes (Anderson, et al., 1996; Joshi et al., 2004).
Coupled with advanced
processing techniques and tools, RS and GIS present a cost-effective, robust, and relatively simple approach to IAP species management, thereby facilitating the decisionmaking process (Underwood et al., 2003). The utility of RS and GIS for mapping of vegetation and IAP species has been well established (Kokaly et al., 2003). The full potential of the technologies, however, acknowledged by Wannenburgh (2009: pers comm.6), is yet to be exploited in South Africa.
This chapter evaluates the utility of 10m multispectral SPOT 5 satellite imagery for the detection and mapping of Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius in KwaZulu-Natal, South Africa. A whole-pixel approach was applied to training and classification of the satellite image data. Specific objectives of this chapter included (1) to test the utility of SPOT 5 medium resolution multispectral imagery for the detection and mapping of the four selected IAP species, and (2) to determine the ‗best‘ algorithm for training and classification. Issues with regards to 6
Wannenburgh, A., personal communication. Assistant director: Decision support, WfW (National office): DWAF
45
resolution, selection of training data, GPS accuracy, and ground-truthing, are also discussed.
3.2
Materials and Methods
3.2.1 Study area and target species The study area is the province of KwaZulu-Natal (Figure 3.1), lying on the eastern seaboard of South Africa between latitudes 26° and 31° S, and longitudes 28° and 32° E. Four study sites were identified, namely, northern KZN (lowland woody veld), coastal KZN (coastal lowlands), KZN midlands (mist belt), and KZN highlands (upland grassland) (Low and Rebelo, 1998).
Preliminary site visits were instrumental in identifying the IAP species at each of the study sites. The four species identified and selected for the study included Acacia mearnsii in the KwaZulu-Natal midlands, Lantana camara in the coastal plains of KwaZulu-Natal, Parthenium hysterophorus in northern KwaZulu-Natal, and Rubus cuneifolius in the KwaZulu-Natal highlands.
46
Figure 3.1: Location of the study area and study sites. Map (A) shows the location of KwaZulu-Natal in South Africa, map (B) shows the four biomes of KwaZulu-Natal (after Low and Rebelo, 1998), map (C) shows the location of the four study sites, map (D) indicates the four SPOT 5 scenes covering the each of the four study sites, and map (E) indicates the scene acquired for the KZN-midlands.
3.2.2 Data acquisition A single orthorectified, radiometrically-corrected (all corrections undertaken by the provider), four-band, 10m multispectral SPOT 5 image covering the KwaZulu-Natal midlands was acquired 17 July 2007. Scenes for each of the remaining three study areas were acquired between 22 and 27 March 2009.
Field campaigns were undertaken to coincide with the time of image acquisition. This was significant as time delays between ground-truthing and image acquisition may have 47
been marked by temporal changes in the environment (e.g. weather and/or changes in species growth, density, or die-off). GPS coordinates of the target species were captured using a Magellan SporTrak handheld GPS receiver, with a recorded horizontal positional accuracy of less than 3m (differentially corrected). Way points were saved to memory as well as recorded in hardcopy for quality assurance during data download, display, and analysis in ArcMap (ESRI, 2008b). This process was significant as coordinates needed to be manually input into the computer (keycoding), from which a database file was created for subsequent processing in ArcMap (ESRI, 2008b). GPS coordinates were used to delineate polygons of species‘ presence data and saved as ESRI shapefiles for subsequent analysis. This procedure was completed for all study areas except the KZN midlands where Sappi forest compartment data was readily available as ESRI shapefiles, providing details regarding the species present in each compartment. These shapefiles were overlaid on the images and polygons containing A. mearnsii subsequently identified and extracted. These polygons, together with those delineated during field campaigns, served as training sites for signature development and ultimately, target detection and classification. Critical to any supervised classification strategy, training data must be ‗both representative and complete‘ (Lillesand et al., 2008: 557), i.e. training data should spectrally define each and all classes in the image (Kavzoglu, 2009). To acquire the ‗best‘ quality training data, the minimum patch size for all targets of interest in the field was established at 20x20m (for SPOT 5) thereby ensuring the sensor adequately captured a whole pixel; the spatial resolutions of SPOT 5 is 10m. This does not, however, indicate the presence of pure pixels, as species‘ ecological and phenological attributes inform both its spatial determinants as well as its density. Species‘ density varied from L. camara forming dense impenetrable clumps to dense, closed-canopy, monospecific stands for A. mearnsii (Figure 3.2). Hence, it was highly improbable that pure pixels be acquired for all targets. For the purpose of this study, pixels with target densities ≥ 80% were accepted as pure pixels (ERDAS, 2008c). To ensure a representative sampling strategy, extensive ground-truthing was undertaken in so far as time, financial, and accessibility 48
constraints would allow.
Ancillary datasets utilised in this study included forest
compartment boundaries sourced from Sappi Forests, and sugarcane plantation boundaries sourced from the South African Sugarcane Research Institute. Both datasets were supplied in ESRI shapefile format for ready integration into a GIS.
Figure 3.2: The target species occurring in the field. Clockwise, from top left: Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius.
3.2.3 Data processing and analysis The aim of this study was to map IAP species by classifying a single class of interest using the image data and several classification decision rules. Foody et al. (2006) assert that a conventional supervised classification would be inappropriate, given the need to 49
exhaustively define each spectral class in spectral space (Congalton and Green, 2009). In this study, a hybrid classification approach (Liu and Mason, 2009) was adopted; an unsupervised (ISODATA) classification was first performed, the results of the classification interpreted using groundtruth data, and subsequently followed by a supervised classification (using four traditional classifiers viz. as Minimum-distance-tomean, Parallelepiped, Mahalanobis and Maximum Likelihood algorithms) using the statistics of the ISODATA classification, as training data. Processing and analysis of the image data involved a several stage process, including pre-processing, whole-pixel training and classification, and post-classification processing.
3.2.3.1 Pre-processing The SPOT 5 datasets were provided as tagged-image files with embedded georeferencing information (*.geoTIFF), in band sequential (BSQ) format.
The images were first
converted to band interleaved by line (BIL) format, subsequently saved to ENVI standard files (ENVI, 2006), and finally imported as ERDAS IMAGINE (*.img) files (ERDAS, 2008a) for ease of analysis, particularly for input to the IMAGINE sub-pixel classification.
To account for the influence of atmospheric disturbances (absorption and scattering), variations in illumination geometry, and fluctuations in viewing angle and solar illumination angle (Yuan and Niu, 2008) on pixel values, radiance values were converted to reflectance values using an atmospheric correction algorithm. The four SPOT 5 scenes were atmospherically corrected using the FLAASH module as implemented in ENVI (ENVI, 2006).
Application of the sub-pixel classifier, as implemented in ERDAS
(ERDAS, 2008c), did not require the scenes be atmospherically corrected, as it contains an included environmental correction module as part of the processing scheme (Table 3.3: 44).
Given a strong degree of inter-band correlation often characteristic of multispectral images (Lillesand et al., 2008; Eastman, 2009b), a principal components analysis (PCA)
50
was applied to the 4-band SPOT 5 image datasets.
Application of the principal
components analysis yielded 4 principal component layers for SPOT 5 (Table 3.1). Evaluating the results of the principal components analysis, the first 3 components for SPOT 5 were input as a data source for training and classification.
A. mearnsii (%)
Eigen values
Table 3.1: Results of the Principal Components Analysis.
L. camara P. hysterophorus R. cuneifolius
Principal Component Layers 1 2 3 4 93.4 4.99 1.31 0.23 90.6 92.7 96.4
7.06 5.31 2.70
2.02 1.81 0.80
0.32 0.15 0.09
3.2.3.2 Whole-pixel training and classification Non-parametric classification algorithms are predominantly utilised when the number of sampled (training) data per class is limited (Richards, 1986). Often, non-parametric classifiers adopt deterministic models in their assignment of pixels to classes, i.e. class assignment is based on some equation that defines pixel allocation to classes. Pixel clusters produced by clustering algorithms (unsupervised training) utilise all/most pixels contained in the input dataset while disregarding contiguity of pixels defining each cluster (ERDAS, 2008a; Lillesand et al., 2008).
The Iterative Self-Organising Data Analysis Technique (ISODATA) (Tou and Gonzalez, 1974) clustering algorithm, which uses the minimum spectral distance formula performed in an iterative process to define spectral classes, was used to identify, isolate, and extract target and non-target spectral classes from the image data. Unsupervised training was significant to the processing schema, as it provided an indication of the number of unique spectral classes in the image, and was used to ‗inform‘ the supervised classification. Furthermore, the non-target class pixels would be later utilised as input data to the maximum likelihood classification (as a non-target signature) and to the IMAGINE 51
Subpixel Classifier classification (as ‗input false aoi‘).
Iterative training of the satellite data using the ISODATA classifier was undertaken to provide a range of output classifications containing N clusters (spectral classes). An initial 8 classifications were produced for each study site; each classification defining N = 50, N = 40, N = 30, N = 25, N = 20, and N = 15, 10, and 5 classes respectively. Each classification was then evaluated against the training data, by overlaying the training sites with the ISODATA-trained image. This process resulted in a match lying between 10 and 15 classes for all sites. Classifications of N = 11, 12, 13, and 14 classes were subsequently run and evaluated against the training sites.
Signature files were edited to best represent the number of spectral classes in the image. This was achieved by combining class signatures based on their mean signature plot (Figure 3.4). Class means that were close to each other, for example class 6 (cyan) and class10 (purple) could readily be combined into a single class, and the ‗new‘ class separability evaluated against the target class using the transformed divergence separability index.
Figure 3.3: Mean signature plot of 10-class ISODATA for SPOT 5.
The target class is
highlighted in green.
52
Class separability was evaluated for each edited ISODATA signature file using the Transformed Divergence separability index (TDij).
Computation of this statistical
distance ensured signatures were distinct from each other, thereby maximising the classification (ERDAS, 2008a). TDij is derived from the computation of divergence (Dij) (Swain and Davis, 1978, Swain and King, 1973):
Dij
1 tr Ci 2
(3.1)
TDij
C j Ci
1
Cj
2000 1 exp
1
1 tr Ci 2
1
Cj
1
T i
j
i
j
Dij 8
(3.2) Where: i and j = the two classes being compared Ci
= the covariance matrix of signature i
µi
= the mean vector of signature i
tr
= the trace function (matrix algebra)
T
= the transposition function
Computation of TDij was crucial to evaluating whether each of the signatures represented a unique class that was spectrally separable from each target class (A. mearnsii, L. camara, P. hysterophorus, and R. cuneifolius) in feature space. For the purpose of this study, only two derived classes; 1 representing the target, and another representing arbitrary non-targets, were required. To achieve this, each of the signature files was iteratively edited to attain a final set of signatures that met this requirement.
The
resulting signature file, for each of the study sites, comprised 5 signatures; 1 signature representing the target class, and 4 signatures representing 4 arbitrary classes that were spectrally separable from the target.
As a general rule of thumb, a TDij value of 1800-2000 indicates separability; a value less than 1800 is indicative of lesser distinct classes, and a value of 0 indicates that the 53
signatures are inseparable. For the purpose of this study, only values for target-nontarget class pairs were of significance; class 1 representing the target species and classes 2, 3, 4, and 5 representing arbitrary spectral classes, separable from the target class (Table 3.2). Table 3.2: Results of the Transformed Divergence (TDij) separability index.
SPOT 5
1:2
Class pairs 1:3 1:4
1:5
A. mearnsii 1862 1963 2000 2000 L. camara 2000 2000 2000 2000 P. hysterophorus 2000 2000 2000 2000 R. cuneifolius
2000 2000 2000 2000
Results of the TDij showed that of the 5 derived classes, the target class (class 1) was spectrally separable from all the other 4 classes. This is evidenced by TDij values of 1800-2000 for all target-non-target class pairs.
Parametric approaches to image classification utilise statistics whereby pixel labelling or classification is executed by taking into consideration statistical parameters such as the mean, standard deviation, and covariance (Schalkoff, 1992). Each pixel value (position) in relation to informational class statistics (mean, standard deviation, and covariance) is then used to determine to which class, pixels are assigned.
The maximum likelihood (ML) classifier is a parametric classifier, based on pattern recognition techniques (Dixon and Candade, 2008), and uses statistical models such as a normal (Gaussian) distribution to define pattern or spread of data (Campbell, 2002; Lillesand et al., 2008; Schalkoff, 1992). The classifier is sensitive to variations in the quality of training data (observed informational classes used to learn the pattern) (Campbell, 2002), and as a consequence, effectiveness of its application depends on a reasonable accurate estimation of the mean vector position and covariance matrix for each spectral class (Richards, 1986).
Accurate estimation of the mean vector and
54
covariance matrix is in turn dependent upon having a sufficient set of training data.
Classification of unknown pixels (constituting a specific vector position) is such that for an observed (training or sampled data) informational class mean value, the ML algorithm computes the probability of unknown pixels falling into one of the observed informational classes where the highest probability determines to which informational class the unknown pixels belong (Campbell, 2002). The ML algorithm assumes equal probability of occurrence and misclassification for all spectral classes (Lillesand et al., 2008).
Training and classification of the image data was undertaken using the ML algorithm as implemented in ERDAS IMAGINE (ERDAS, 2008b) and calculated using:
D ln ac
0.5 ln Covc
0.5 X M c T Covc 1 X M c
Covc
(3.3)
Where: D
= weighted distance (likelihood)
c
= the target class
X
= the measurement vector of the candidate pixel
Mc
= the determinant of the covariance matrix of the data in class
ac
= percent probability that a candidate pixel is a member of class c
i
Covc = the covariance matrix of the pixels in the sample of class c
Covc = determinant of Covc (matrix algebra) Covc 1
= inverse of Covc (matrix algebra)
ln
= natural logarithm function
T
= transposition function (matrix algebra)
3.2.3.3 Post-classification processing 3.2.3.3.1 Post-classification smoothing The salt-and-pepper appearance often characteristic of classified data (Lillesand et al., 55
2008, Lu and Weng, 2007), was ‗smoothed‘ of this small erroneous data thereby reducing the (irrelevant) detail for a more general analysis. The ‗Majority Filter‘ was applied to the image as a 3x3 matrix kernel (moving window) assigning a new pixel value to the misclassified pixels based on the majority of their contiguous neighbouring pixels (Lillesand et al., 2008; ESRI, 2008a). The application of a 3x3 matrix kernel was justified by the location accuracy (30m) of the satellite imagery coupled with the positional accuracy of the GPS receiver (3m). Pixel replacement was conditional upon a satisfactory number of similar-value neighbouring pixels, and contiguity of these neighbouring pixels about the centre of the kernel (ESRI, 2008a).
3.2.3.3.2 Classification accuracy assessment ‗A classification is not complete until its accuracy is assessed‘ (Lillesand et al., 2008: 585). Several principles and practices are currently used for assessing classification accuracy; the commonest being the error matrix, which simply assesses the relationship between groundtruth data and the results of the classification algorithm (Congalton and Green, 2009; Lillesand et al., 2008).
Accuracy assessments are indispensable for reporting on the classification through identification and correction of error, facilitating comparison of the various algorithms and techniques used, and for determining the relevance of classification products in relation to the decision-making process (Congalton and Green, 2009). In this study, classification accuracy was assessed and reported for target classification as overall accuracy (computed by dividing the sum of correctly classified pixels by the total number of pixels), user’s accuracy (derived by dividing the number of correctly classified pixels in a class by the total number of pixels classified in that class), producer’s accuracy (derived by dividing the number of correctly classified pixels in a class by the total number of reference pixels), and Kappa statistic (a measure of the difference between actual agreement between reference data and an automated classifier, and the chance agreement between the reference data and a random classifier).
To account for the locational uncertainty of the reference points, attributed to location 56
accuracy of the imagery and positional accuracy of the GPS receiver, a ‗conservative approach‘ was adopted in determining the classification value for each reference (target and non-target) point. As opposed to taking the value of the pixel corresponding to the reference point, the value was defined as the majority class within a specified distance (20m) from the reference point (Jenness and Wynne, 2007a). The classification value is hence the majority value in a circular area about that point; analogous to generating a buffer around each point at a radial distance of 20m and calculating the average value of pixels within the buffer zone.
The testing data comprised reference points captured in the field using a handheld GPS receiver, and appended to randomly-generated reference points created using Hawths Analysis Tools (Beyer, 2006). This provided a compromise between a purely random, and a ‗field-random‘ sample. A minimum of 100 reference points were used to calculate the accuracy statistics, which was undertaken and evaluated in ArcView 3.3 (ESRI, 2002), using Kappa Tools (Jenness and Wynne, 2007b).
3.3Results and Discussion 3.3.1 Unsupervised classification The first step in applying a binary classification was to run an unsupervised classification on the image data. The ISODATA algorithm, as noted in the methodology, was used as an ‗indicator‘ of the most likely number of unique spectral classes in the image, and was subsequently used to ‗inform‘ the application of the supervised classification algorithms; the second step of the binary classification. Results of the ISODATA classification (Figure 3.4) was significant in providing an indication of the spatial distribution and location of clusters (unique spectral classes), and coupled with groundtruth data, facilitated the digitising of areas of interest (*.aoi) for supervised training.
Initial supervised training highlighted the need to provide the algorithm with more training data than was acquired in the field. It thus became necessary to acquire training data for ‗other‘ (non-target) classes to enable the algorithm to effectively discriminate 57
and differentiate the data into unique spectral classes. This was achieved by several means, viz. using elements of visual image interpretation (shape, pattern, and texture) to identify objects (e.g. urban areas), through the use of cadastral data (e.g. boundaries delineating agricultural land parcels), level-slicing (e.g. to ‗highlight‘ waterbodies), and using NDVI (as an indication of vegetation biomass). This exercise was coupled with the results achieved using the ISODATA classifier. Training areas were isolated to provide at least one hundred (non-target) pixels for classification.
Figure 3.4: Unsupervised training of SPOT 5 using the ISODATA algorithm. Map (A) Acacia mearnsii, map (B) Lantana camara, map (C) Parthenium hysterophorus, and map (D) Rubus cuneifolius. Inserts show overlay with testing data (training sites).
58
3.3.2 Supervised classification The results presented in the following discussion were used to inform the selection of classifiers to be used in the mapping protocol of the alien plant species under consideration.
In each study site four classifier algorithms were used.
The
classification performance of Parallelepiped, Mahalanobis-Distance, MaximumLikelihood and Minimum-Distance-to-Mean classifiers here onwards referred to as BOX, MAHAL, ML and MDM respectively were assessed to determine the accuracy of each classifier.
3.3.2.1 Classification accuracy assessment: A. meanrsii distribution maps Table 3.3 (page 60) is a summary accuracy report generated from error matrices of the respective classifier algorithms chosen. In mapping the distribution of A. mearnsii (Figure 3.5:62), 0.8 (Kappa: 0.6), 0.63 (Kappa: 0.25) and 0.63 (Kappa 0.25) overall accuracies were produced for the ML, MDM and MAHAL classifier algorithms respectively. A close visual examination of the A. mearnsii distribution maps (Figure 3.5) generated from ML, MD and MAHAL classifier algorithms show slight variations with regards to A. mearnsii presence.
The accuracy report for the BOX classifier algorithm was invalid since there were no pixels classified as A. mearnsii at the observed or ground reference locations used to assess the accuracy of the classification. The BOX classifier was not able to identify A. mearnsii at any of the observed or ground reference locations i.e. pixels at observed or ground reference locations were left unclassified. This can partly be explained by class assignment decision rule of the BOX classifier i.e. for a BOX classifier, all pixels outside of the range or ‗box‘ remain unclassified and therefore the user‘s accuracy report was nullified (Table 3.3). The producer‘s accuracies for both MDM and MAHAL indicate that only 0.60 of the observed or ground reference locations known to be A. mearnsii were correctly classified as A. mearnsii (Table 3.3). On the other hand, the user‘s accuracies for both 59
MDM and MAHAL indicate that only 0.59 observed or ground reference locations of the classified as A. mearnsii were indeed known to be A. mearnsii (Table 3.3).
Table 3.3 is a classification performance summary report of error matrices from the respective classifier algorithms. Note that the producer‘s and user‘s accuracy is specific to target class only. KZN midlands A. mearnsii Classifier algorithm
ML
MDM
MAHAL
BOX
Overall accuracy
0.80
0.63
0.63
0.55
User‘s accuracy
0.78
0.59
0.59
Null
Producer‘s accuracy
0.78
0.60
0.60
0.00
Kappa
0.60
0.25
0.25
0.00
coefficient
(K-Hat)
The user‘s and producer‘s accuracies of 0.78 for the ML classifier algorithm indicate to a good mapping of A. mearnsii distribution (Table 3.3). An overall accuracy and Kappa for ML (0.80 and 0.60 respectively) was found to be higher than that of MDM and MAHAL (Table 3.3). Noticeably, the errors of commission and omission (0.36 and 0.13) for ML were considerably lower than that of MDM and MAHAL (Table 3.4).
Table 3.4: Commission and omission errors from the respective classifier algorithms generated from the image covering KZN midlands. Note that the errors of commission and omission are specific to A. mearnsii only. SPOT 5
ML MDM MAHAL BOX
Omission 0.36 0.67 0.45 Null
Commission 0.13 0.56 0.76 Null
Thus, the ML classifier algorithm performance suggests that the distribution of A. mearnsii (Figure 3.5) is best mapped using SPOT 5 and the ML decision rule since it 60
produced the highest users‘ accuracy (Table 3.3), coupled with a lowest error of commission (Table 3.4).
61
Figure 3.5: Shows the distribution of A. mearnsii using ML, MDM, MAHAL and BOX classifier algorithms
62
3.3.2.2 Classification accuracy assessment: R. cuneifolius distribution maps Table 3.5 (page 64) is a summation of the accuracy reports generated from error matrices of the respective classifier algorithms. Overall accuracies of 0.60 (Kappa: 0.19), 0.34 (Kappa: 0.00) and 0.47 (Kappa 0.03) were obtained using the ML, MDM and MAHAL classifier algorithms correspondingly in order to map R. cuneifolius. A close visual assessment of the distribution maps for R. cuneifolius (Figure 3.6: 65) reveals that the BOX classifier assigned no pixels to R. cunefolius whilst MDM and MAHAL assigned unrealistic pixels to R. cuneifolius .
In the case of BOX classifier algorithm, the absence of pixels classified as R. cunefolius may also be related to class assignment decision rule i.e. for an unknown pixel that lies in the intersection of classes will be assigned to a class the BOX classifier algorithm encounters first. Furthermore, the class assignment decision rule is such that, all pixel values falling outside of the range or ‗box‘ remain unclassified. Evidently, this was the case with observed or ground reference locations where R. cunefolius is known to be present. Thus in the case of mapping R. cunefolius using the BOX classifier, the BOX classifier the accuracy report was nullified due to inability to identify R. cunefolius at the observed or ground referenced location (Table 3.5). The producer‘s accuracies of the MDM and MAHAL classifiers indicate that 0.04 and 0.38 of the observed or ground reference locations known to be R. cunefolius were correctly mapped respectively (Table 3.5). Whilst the user‘s accuracies for both MDM and MAHAL indicates respectively that at least 0.71 and 0.70 of the observed or ground reference locations classified as R. cunefolius were indeed known to be R. cunefolius (Table 3.5). Visual evaluations of the classification results generated from MDM and MAHAL showed that classification resulted in unrealistic presence of R. cunefolius. This can partly be attributed to the system confusing other vegetation types with R. cunefolius. The user‘s accuracy of 0.77 and producer‘s accuracy of 0.59 for the ML classifier
63
algorithm indicates a good mapping of R. cunefolius (Table 3.5). An overall accuracy and Kappa for ML (0.60 and 0.19 respectively) was found to be higher than that of MDM and MAHAL (Table 3.5). Clearly, the errors of commission and omission (0.17 and 0.56 respectively) for ML were considerably lower than that of MDM and MAHAL (Table 3.5). Hence, the ML classifier algorithm performance suggests that the distribution R. cunefolius is best mapped using SPOT 5 and the ML decision rule since it produced the highest users‘ and producer‘s accuracies(Table 3.5), coupled with a lowest errors of commission (Table 3.6). Table 3.5 is a classification performance summary report of BOX, MAHAL, ML and MDM. Note that the producer‘s and user‘s accuracy is specific to R. cunefolius only. KZN highlands R. cunefolius Classifier algorithm
ML
MDM
MAHAL
BOX
Overall accuracy
0.60
0.34
0.47
0.32
User‘s accuracy
0.77
0.71
0.70
Null
Producer‘s accuracy
0.59
0.04
0.38
0.000
Kappa
0.19
0.00
0.03
0.000
coefficient
(K-Hat)
Table 3.6: Commission and omission error for ML, MDM, MAHAL and BOX algorithm.
Note that the errors of commission and omission are specific to R.
cunefolius only.
SPOT 5
ML MDM MAHAL BOX
Omission 0.56 0.63 0.69
Commission 0.17 0.45 0.68
64
Figure 3.6: Shows distribution of target species R. cunefolius in KZN highlands obtained using MAHAL, MDM, BOX and ML classifiers
65
3.3.2.3 Classification accuracy assessment: P. hysterophorus distribution maps A summation of the accuracy reports generated from error matrices of the respective classifier algorithms is shown in Table 3.7(page 66). The classification of SPOT 5 data to map P. hysterophorus produced overall accuracies of 0.54 (Kappa: 0.06), 0.50 (Kappa: 0.04) and 0.41 (Kappa 0.01) for the MDM, MDM and AHAL respectively. A close visual examination of the P. hysterophorus distribution maps (Figure 3.7: 67) generated from the classification shows that the BOX, MDM and MAHAL classifiers assigned unrealistic pixels to P. hysterophorus. However, the Box classifier was not able to identify P. hysterophorus at the observed or ground reference locations and thus, the user‘s accuracy report was nullified (Table 3.7).
The over-exaggerated presence of P. hysterophorus in the case of MDM, MAHAL and BOX classifiers is indicative of classification system confusing other vegetation types such grass to be P. hysterophorus. As noted in the mapping of R. cunefolius and A. mearnsii using the BOX classifier algorithm, the unclassified pixels at observed or ground reference locations may be attributed to pixel values falling outside of the range or ‗box‘ that determines class assignment.
Table 3.7is a classification performance summary report of BOX, MAHAL, ML and MDM. Note that the producer‘s and user‘s accuracy is specific to P. hysterophorus only. Northern KZN P. hysterophorus Classifier algorithm
ML
MDM
MAHAL
BOX
Overall accuracy
0.54
0.50
0.41
0.40
User‘s accuracy
0.63
0.62
0.60
Null
Producer‘s accuracy
0.56
0.46
0.01
0.000
Kappa
0.06
0.04
0.01
0.000
coefficient
(K-Hat)
66
Figure 3.7: Shows the distribution of P. hysterophorus in Northern KZN obtained using MAHAL, MDM, BOX and ML classifiers
67
The producer‘s accuracies for MAHAL and MDM classifiers indicate respectively that 0.46 and 0.01 of the observed or ground reference locations known to be P. hysterophorus were correctly mapped (Table3.7). Results of the user‘s accuracies for the MDM and MAHAL on the other hand, indicate respectively that 0.62 and 0.60 of the observed or ground reference locations classified as P. hysterophorus, were indeed known to be P. hysterophorus (Table 3.7). Table 3.8: Commission and omission error for ML, MDM, MAHAL and BOX algorithm.
Note that the errors of commission and omission are specific to P.
hysterophorus only. SPOT 5
ML MDM MAHAL BOX
Omission 0.31 0.78 0.60 Null
Commission 0.20 0.58 0.69 Null
The user‘s accuracy of 0.63 and producer‘s accuracy of 0.56 for the ML classifier algorithm indicates a good mapping of the distribution of P. hysterophorus (Table 3.7). An overall accuracy and Kappa for ML (0.54 and 0.06 respectively) was found to be higher than that of MDM and MAHAL (Table 3.7). The errors of commission and omission (0.31 and 0.20 respectively) for ML were considerably lower than that of MDM and MAHAL (Table 3.8). Hence, the ML classifier algorithm performance suggests that the distribution P. hysterophorus is best mapped using SPOT 5 and the ML decision rule since it produced the highest users‘ and producer‘s accuracies, coupled with a lowest errors of commission.
3.3.2.4 Classification accuracy assessment: L. camara distribution maps
Table 3.9 (page 69) shows a summation of the accuracy reports generated from error matrices of the respective classifier algorithms. A close assessment of the distribution maps for L. camara (Figure 3.8: 71) generated from MDM MAHAL and BOX classifiers shows that, the BOX, MDM and MAHAL classifier assigned unrealistic pixels to L. camara. However, the BOX, MDM and MAHAL classifiers were unable 68
identified L. camara at the observed or ground reference locations and thus the user‘s accuracies were invalid (Table 3.9).
It has been noted in the mapping of R.
cunefolius, A. mearnsii and P. hysterophorus distribution using the BOX classifier, that the unclassified pixels at the observed or ground reference locations may be attributed to pixel values falling outside of the range or ‗box‘ that determines class assignment. In the case of MDM and MAHAL classifiers, the unclassified pixels can be attributed to fact that, the MDM and MAHAL classification modules were set up to leave pixel in the intersection of two or more spectral classes as unclassified. However over-exaggerated presence of L. camara in the case of MDM, MAHAL and BOX classifiers is indicative of classification system confusing other vegetation types such Chromolaena Odorata (paraffin weed) and other shrubs to be L. camara.
Table 3.9 is a classification performance summary report of BOX, MAHAL, ML and DM. Note that the producer‘s and user‘s accuracy is specific to L. camara only. Coastal KZN L. camara Classifier algorithm
ML
MDM
MAHAL
BOX
Overall accuracy
0.74
0.79
0.79
0.79
User‘s accuracy
0.41
Null
Null
Null
Producer‘s accuracy
0.44
0.000
0.000
0.000
Kappa
0.26
0.000
0.000
0.000
coefficient
(K-Hat)
The user‘s accuracy of 0.41 and producer‘s accuracy of 0.44 for the ML classifier algorithm indicates a modest mapping of the distribution of L. camara (Table 3.9). An overall accuracy and Kappa for ML was found to be 0.74 and 0.26 respectively (Table 3.9). Errors of commission and omission were found to be 0.90 and 0.06 respectively for the ML classifier (Table 3.10: 70). The ML classifier algorithm performance suggests that the distribution L. camara is best mapped using SPOT 5 and the ML decision rule since accuracy reports of MDM, MAHAL and BOX were all nullified.
69
Table 3.10: Commission and omission error for ML, MDM, MAHAL and BOX algorithm. Note that the producer‘s and user‘s accuracy is specific to L. camara only. SPOT 5
ML MDM MAHAL BOX
Omission 0.90 Null Null Null
Commission 0.06 Null Null Null
70
Figure 3.8: Shows the distribution of L. camara in Coastal KZN obtained using MAHAL, MDM, BOX and ML classifiers
71
3.3 3 Scale, Resolution, and IAP Species Mapping Foxcroft et al. (2009) recorded several important patterns and observations with regards to mapping IAP species at different spatial scales (quaternary catchment to 250m), spatial scale having concomitant monitoring and management implications. The scale of mapping is particularly significant with regards to estimating IAP species richness, identifying spatial structure and distribution, and understanding spread dynamics and ecology. For remotely sensed mapping of IAP species, image scale is a significant parameter to selecting appropriate sensor(s); broad-scale vegetation mapping may be undertaken using coarse resolution sensors (e.g. MODIS: moderate resolution imaging spectroradiometer), but higher resolution sensors (e.g. SPOT) is more appropriate for landscape-scale vegetation mapping (Madden et al., 2004), with parallel effects in relation to the accuracy of information extraction.
Vegetation patch size is significantly important in sensor selection, given the varied spatial resolution of satellite imagery, as well as the inconsistent spatial extents (and densities) of the target invader species.
Sensor resolution is thus an important
consideration in mapping vegetation, as both spatial and spectral resolution impact on the classification accuracy (Lawrence et al., 2006).
In this study, 10m SPOT 5
imagery required an ideal patch size of 20m, ensuring that a pixel completely delineated the ground target (IAP species patch). In many cases this was not feasible given landscape fragmentation, IAP species patchiness, and IAP species‘ association with grass and other vegetation, resulting in mixed pixels containing several classes (spectral signatures).
3.4
Conclusions
Traditional classifiers (e.g. maximum likelihood) are based upon the user defining pure pixels (≥80% density) to acquire a satisfactory classification of a target material; ML mapped A. mearnsii with an overall accuracy of 0.8 (Kappa: 0.6). This is not always possible, particularly in fragmented landscapes and cases of high spectral 72
mixing (mixed pixels). The adoption of ‗soft‘ classifiers (e.g. neural networks) would be more appropriate in such cases. The attempt to map IAP species in the heterogeneous landscape of KwaZulu-Natal has attained more than satisfactory results. The utility of SPOT 5 medium resolution multispectral data incorporating the maximum likelihood classification algorithm, has been shown to be effective in mapping A. mearnsii in the KwaZulu-Natal midlands, and L. camara in the coastal plains of KwaZulu-Natal, less effective at mapping R. cuneifolius in the KwaZulu-Natal highlands, and ineffective in mapping P. hysterophorus in the northern plains of KwaZulu-Natal.
Although the influence of grass and other vegetation may have contributed to the low overall classification accuracy, more significant is the difficulties of classifying P. hysterphorus and R. cuneifolius on the basis of SPOT 5 data acquired during autumn/early winter. Phenological changes, exhibited as winter die-off, make for extremely difficult detection. This was further made difficult by the spatial and spectral resolution of SPOT 5 data. It is thus essential that future research take cognisance of species phenology and plan image acquisition accordingly, although this may not be feasible for planned image acquisition of large spatial coverages.
73
Chapter 4: Semi-Automated Mapping 4.1
Introduction
This chapter briefly highlighting the potential offered by integrating remotely sensed data and GIS. Examples of platforms that integrate GIS and remote sensing data are provided, in order to, demonstrate the various ways in which image processing tailored for a specific task can be created and automated. Particular attention is given to contrasting customised applications developed using script-based programming and those that employ visual-based programming. The discussion draws to a conclusion by outlining the motivation for the study and by providing the aim and objectives for the study.
Remote sensing is driven by the need to acquire information about earth resources (Sharifi, 2002, Strahler et al., 1986). Consequently, remotely sensed data such as multispectral (XS) airborne and satellite imagery are increasingly available for detection, delineation, and classification of vegetation (Bobbe et al., 2001). Madden (2004) sees the current use of GIS and remote sensing to map vegetation expanding considerably, due to public demand for a better environmental protection on the one hand, whilst on the other hand suggests that, improvements in satellite imagery resolution, GPS, computers and software are also spurring the application of these technologies.
Clearly, it is to be expected that resource managers would find
themselves having to process large amounts of data prior to sanctioning action. More importantly, there is a clear necessity to act swiftly on the basis of information gathered when appraising AIP species.
In order to address some of the challenges that may arise with handling large volumes of data, questions around human resource and expertise in processing data need to be addressed. The inability to provide thematic information at the required speed, quality and cost remains an obstacle within the remote sensing community (Luck and Zietsman, 2006). However, the ability to combine GIS and remote sensing techniques renders a powerful tool able to quickly map, and model the distribution of vegetation types (Madden, 2004). Essentially, such a tool requires integration of remotely sensed 74
data into a GIS that has image processing functionality, such that GIS projects with remotely sensed data can be readily converted to thematic maps.
4.2
Automating Image Processing: Platforms used to Create
Protocols and Algorithms Broadly protocols can be defined as a set of procedures or rules carried out to achieve any desired outcome (Nixon and Aguado, 2000). A simple way of doing this is by drawing symbols that represent all the components involved and the processes operated on them (Eastman, 2006). A computer program is an all encompassing example of an algorithm (Nixon and Aguado, 2000). In its most general sense, an algorithm is somewhat similar to a protocol since it also details instructions which result in a desired outcome.
This is primarily due to algorithms (or computer
programs) being simply a series of instructions of varying degree of complexity (Nixon and Aguado, 2000). However, algorithms in computers execute a set of procedures or instructions automatically.
Basically procedures are automated
primarily because it may be difficult to keep track of the various datasets, processing procedures, parameters and assumptions used during analysis (ESRI, 2000). Thus automation of image classification offers the potential to produce maps faster and at consistent accuracy levels More importantly, in ERDAS for example this allows analyst to process one or more files with one or more commands at any time, from one minute to many years in the future (Leica Geosystems, 2006).
Currently image processing and computer vision systems can be implemented in various system and scripting languages (Nixon and Aguado, 2000; ESRI, 2006). System languages such as C++ and Java TM are used to create applications from scratch using raw resources of the computer, whilst scripting languages such as Python on the other hand are used to glue applications together (ESRI, 2006). These languages enable users to create customized tools and procedures used in the development and execution of various applications (Nixon and Aguado, 2000).
However, image
processing executed using scripting languages requires competency in the use of the 75
language syntax. Hence this limits or confines the user community to that of expert users.
The introduction of software that combines remotely sensed data and GIS coupled with image processing functionality not only provides users with in-built tools but also, provides users with the possibility of creating customized procedures for image processing. ERDAS Imagine 9.3 and ArcGIS-ArcEditor 9.3 are some of the commercial software packages that provide image processing toolkit coupled with the possibility of creating user defined procedures for applications.
The creation of
customized procedures is not only operated using language syntax but also, userfriendly visual programming language also known as graphical modelling.
The
graphical models can be saved and used for automating procedures or creating algorithms (ESRI, 2000; Leica Geosystems, 2006; Gonzalez et al., 2004). Basically procedures are automated primarily because it may be difficult to keep track of the various datasets, processing procedures, parameters and assumptions used during analysis (ESRI, 2000). More importantly, software packages like ERDAS Imagine 9.3 for example, allow analysts to process one or more files with one or more commands (Leica Geosystems, 2006). As mentioned above, instruction compilation approaches can either be based on scripts (text or code) or symbols (graphics) process models. In ERDAS Imagine 9.3 for example, Spatial Modeller Language (SML) interface (Figure 4.1) is used to develop code (script) based models (Leica Geosystems, 2006). Using this interface, process models can be created such that they best suit the data and the objectives.
76
Figure 4.1: Portion of a script-based algorithm that runs via SML background. The algorithm converts digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map A. mearnsii. Alternatively, the Model Maker interface shown in Figure 5.2 can be used to execute the same functions used in the script based model. Essentially, the symbols employed in Model Maker are a graphical version of SML scripting interface. The Model Maker interface is built in such a way that it runs at the background of SML language where symbols representing functions can be used to create and run models for image processing (Leica Geosystems, 2005). This is particularly useful for users with limited or no language programming skills.
77
Figure 4.2: A graphic based algorithm modelled in Model Maker. The algorithm converts digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map A. mearnsii. In ArcGIS-ArcEditor 9.3, VBscript, Python and Jscript programming interface can be used to develop code or script based models (ESRI, 2006). Like in ERDAS Imagine 9.3, using VBscript interface (Figure 4.3) for example, process models are created such that they best suit the data and the objectives.
However, competency in
programming language is required. Again if there are language competency issues, the ModelBuilder interface (Figure 4.4), which is essentially a graphical version of script based programming interface can be used.
In evaluating available platforms for
creating customised applications, the graphical modelling is clearly the platform of choice. This is primarily because of the fact that workflows describing functions and processes are simplistic, better understood, relatively easy to use and efficient. On the other hand, code based platforms require intermediate to in-depth knowledge of 78
programming language. Thus, visual-based programming was used.
Figure 4.3: Portion of a script-based based algorithm written in VBscript.
The
algorithm converts digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map Acacia. mearnsii.
Figure 4.4: A graphic-based algorithm modelled in ModelBuilder. The algorithm 79
convert digital numbers to radiance values for a SPOT 5 image. The example shown is for KZN midlands image used to map Acacia. mearnsii.
80
4.3
Motivation, Aim and Objectives
In line with the growing need for information about earth resources, the South African government and SPOT-Image, a French Space Agency public company entered into an agreement to capture SPOT 5 imagery for the whole country.
At the time of
compiling this report, government departments and parastatals companies had access to June-July SPOT 5 imagery captured over a three year period. The data volumes will gradual increase over the years and clearly there exist a need to develop methodologies to analyse the data. Secondly, the data sets will have to be processed at relatively short periods of time. It is to the technological advancements outlined in the preceding discussion, coupled with the need for timely produced spatial information that, a component of the overall study aim is to
To test the utility of using SPOT 5 XS imagery to develop algorithms for mapping some of the major current and emergent invasive alien plant species. The objectives of the study were,
1. To specify or write procedures for mapping IAP using the classification of the
highest accuracy
2. To automate the procedures into algorithms
4.4
Study Area
81
4.5
Material and Methods
4.5.1
Data collection: remotely sensed imagery and Global Positioning Systems data
Alien plant species considered for the study occur in different localities within the study area. As a consequence, the researchers decided to map a single alien plant species in a given study site. SPOT 5 multispectral images were employed across all four study sites within KwaZulu-Natal. The image data used were captured on the 17th July 2007 for the KZN midlands study site and between the 22nd and 27th of March 2009 for the Coastal KZN, Northern KZN and KZN highlands study sites . ERDAS Imagine 9.3 and ArcGIS-ArcEditor 9.3 are commercial software packages tested for the development of a customised application used to automatically classify SPOT 5 images. These commercial packages were considered since they provide users with the possibility of creating user-defined procedures or protocols. applications. Lastly and more importantly, the tested software for protocol and algorithm development was determined in part by access and availability on the part of the researchers and the stake holder.
4.5.2 Designing and automating an image analysis scheme: protocol and algorithm development In developing protocols, the desired outcome is to produce a sequence of operations executed for the successful mapping of the target species. To achieve this, different classifier algorithms were used (see Chapter 4) and the classifier that yielded the highest satisfactory results was considered for the final protocol development. The classification results in Chapter 4, indicate that the Maximum Likelihood performed better than the other classifier algorithms tested and thus, it was included in the final protocol. The procedures used to map the chosen IAP species in Chapter 4 were developed into a protocol that was graphically modelled and executed as a semiautomated algorithm for mapping the selected IAP species in ArcGIS-ArcEditor ModelBuilder. The algorithm was validated by classifying an image captured on 17th July 2007. The image covered forested areas in KZN midlands. These forested areas 82
were not included in the images used to test the performance of the Maximum Likelihood, Minimum-distance-to-mean and Mahalanobis classifier algorithms traditional classifiers. Acacia mearnsii was chosen for validating the performance of the semi-automated algorithm for mapping the selected IAP species.
4.6
Results and Discussions
The protocol developed for mapping IAP species assumes that necessary preprocessing of data such as atmospheric correction, geometric correction steps has been performed. Secondly, generating polygons delineating the location of most if not all non-targets class within the scene is paramount. The protocol (Table 5.1) is presented in a series of sequential steps and only three data sets are required i.e. image data, training information and ground-truth or reference point data.
Although data
acquisition is part of the first step, it however, is a step that should be executed prior to using the protocol.
83
Table 4.1 is a lists and describing procedures executed in the successful mapping of selected IAP species Algorithm steps Step 1
Description of model elements
Step 2 Step 3
Step 5
Generate Principal Components Three Principal Components image outputs & Principal Component statistics report (text file) Load training data delineating only one target and a range of nontarget classes Generates Signatures
Step 6
Create training data signatures output file
Step 7 Step 8
Classify using Maximum Likelihood classifier Generate classified image output
Step 9
Perform smoothing using Majority Filter 3x3 window
Step 10
Generate filtered image output
Step 11 Step 12
Reclassify thematic image: 1 for Target and 2 for all Non-Target classes Generate Reclassified image output
Step 13
Load ground-truth point data used to sample image class
Step 14
Sample Image Classes
Step 15
Image class sample points tabulation
Step 4
Load Image Data:
4.6.1 Automation in ArcGIS-ArcEditor 9.3: A semi-automatic algorithm for Mapping selected IAP species In its most general sense, an algorithm is similar to a protocol since it also details instructions which results in a desired outcome. This is primarily due to algorithms (or computer programs) being simply a series of instructions of varying degree of complexity (Nixon and Aguado, 2000). However, algorithms in computers execute a set of procedures or instructions automatically. In ERDAS imagine 9.3, the use of Model Maker to create an algorithm for mapping alien plant species was limited by unavailability of classifiers. It essentially requires a developer‘s toolkit special license to enable compilation of instructions that include classifiers (Leica Geosystems, 2005). 84
Hence development of algorithms using ERDAS was discarded. ArcGIS-ArcEditor 9.3 is a GIS software package with image processing functionality. Not only does this software packages offer in built for modules image process, but it also provide the analyst with opportunity to develop user-defined applications. For the purpose of this research, ModelBuilder in ArcGIS-ArcEditor 9.3 was used since it provides both image processing toolkit and the possibility of the creating customised applications. Elements, connectors and text labels are the three basic components used in ModelBuilder (Figure 5.5) modelling environment (ESRI, 2006). Elements are made up of processing tools, input and derived data. Processing tools are modules used to perform specific tasks or functions on projects‘ input data thereby producing derived data. Consistency with the use of colour and shape to denote processing tools, input and derived data is maintained. A model that is ready to perform specific tasks will have the colour coding shown in Figure 5.5 below for the respective elements of the model. The colours are indicative of the model requirements and instructions properly defined.
Contrary, if a model or a tool is not to be run or not ready to run the fill
colour for the elements will be white.
Figure 4.5: is a schematic showing the basic components of used to build a model in ModelBuilder interface. Elements colours are used symbolise input data, geoprocessing tool and derived data is kept throughout the model. 85
The procedures specified in the protocol were modeled and automated in ModelBuilder.
The geoprocessing tool elements used to execute automatic
classification of were taken from ArcGIS Spatial Analyst toolbox. Table 5.2 lists and describes steps in the algorithm whilst Figure 5.6 shows the image analysis scheme model created and automated within Modelbuilder.
86
Figure 4.6: A schematic representation of an algorithm generated and executed in ModelBuilder. Note colour code representing input elements, geoprocessing tools and derived data.
87
In validating three basic elements for the algorithm were required. The first element was training information which was obtained from the forest-sector detailing forests compartment data. The second was an image covering KZN-midlands. The third and last element was the random points generated within different known compartments. These elements were loaded in the algorithm shown in Figure 5.6 above. The automated classification executed within the ModelBuilder interface was used to map A. mearnsii and the results are shown in Table 5.3 below.
Table 4.2 Algorithm validation report Classifier algorithm
ML
Overall accuracy
0.719
User‘s accuracy Producer‘s accuracy Kappa coefficient (K-Hat)
0.878 0.710 0.246
Figure 4.7: An application designed to run the algorithm 88
Considerations given to the necessary skills and knowledge required to operate ModelBuilder interface lead to development of an application window in Figure 5.7 (page 88). The application tool interface requires three basic data sets i.e. image data,training information and ground-truth or reference point data., which runs in ArcView-ArcMap. At the time of writing the report the application was ran at fixed file locations determined by the developer. Modification of would be effected for other users general users.
4.7
Conclusion
The Maximum Likelihood performed exceptional for A. mearnsii and L. camara however, the performance of Maximum Likelihood in mapping mapping P hysterophorus and R. cunefolius was not the best. In most cases the over exaggeration of species presence can be partly attributed to the fact there exist classes in the each scene that were accounted for in the seperability. This places a huge burden on the efficacy of the algorithm. In using the algorithm all possible classes within the scene should be exhausted. Time and human resource should be invested in getting exceptionally good quality training data. By doing this, the classification procedure will be less likely to commit errors commission. One way getting around impediments of cost and time is to not use the whole scene. It is easier for Maximum Likelihood classifier and other statistical approaches handle distinct class, thus if the classification scheme was only applied to sub sets of scenes, it would be possible for the scheme to perform even better especially for species like P. hysterophorus and R. cunefolius.
89
Chapter 5: Modelling the Range and Distribution of IAPS 5.1
Introduction
An IAP species‘ temporal and spatial distribution pattern is determined by the complex interactions of three principal factors: (1) the species‘ dispersal capacity, (2) the spatial distribution of environmental variables of physiological significance (e.g. temperature), and (3) characteristics of the biotic environment (e.g. competition) (Soberón, 2007). Delineating a species‘ geographic distribution may however be achieved through the modelling of climate-vegetation relationships, based on ecological theory, increased data capacity, and new statistical techniques (Torres-Meza et al., 2009).
The ability to model the range and distribution of IAP species across landscapes has significant implications to their management and control (Anderson et al., 2003; Boylen et al., 2006).
A Geographic information system allows for the development of
multivariate spatially explicit models, incorporating digital environmental data, for delineating species‘ niche habitats (Browning et al., 2005; Hirzel, et al., 2002; Rotenberry et al., 2006). The latest analytical GIS tools coupled with multivariate statistics permit the generation of complex multivariate models of a species‘ distribution at the landscape level (Clark et al., 1993).
A number of modelling algorithms are currently available; their modelling approaches distinguished into 2 types: (1) methods that incorporate both presence and (pseudo)absence data (e.g. MAXENT (maximum entropy) and OM-GARP (open modeller-genetic algorithm for rule-set prediction)), and (2) methods that incorporate presence-only data (e.g. BIOCLIM and DOMAIN) (Giovanelli et al., 2010; Tsoar et al., 2007). Fundamentally, the basis of the modelling approach is determining the species climate envelope by matching climatic and other environmental factors with distribution data (Aurambout et al., 2009).
This chapter evaluates the application of the Mahalanobis distance statistic to predicting species‘ habitat suitability (potential niche), based on climatic and physical environmental parameters defining species‘ distribution, and using species‘ presence90
only data. The Mahalanobis distance (MD) approach utilises the multivariate mean and covariance matrix in defining an oblique elliptic envelope (Figure 5.1) by ranking potential habitat sites by their Mahalanobis distance to a mean vector describing the environmental conditions of all sites in the environmental space (Farber and Kadmon, 2003). The implementation of Mahalanobis distances is explored using the Mahalanobis distances extension for ArcView 3.x (Jenness, 2003b). Analysis was carried out in ArcEditor 9.3 and ArcView 3.3, with the Spatial Analyst extension loaded.
Recorded observations
Compute : Comput eMean vector (size N) and a correlation matrix (size N x N) Find : Climatic combinations within a given Mahalanobis distance to the mean vector
Create : Prediction map by projecting the climatic combinations Figure 5.1: Conceptual framework for the Mahalanobian model (adapted from Farber and Kadmon, 2003).
91
5.2
Materials and Methods
5.2.1 Study area and target species Four bioclimatic regions (Phillips, 1973) were identified within the study area of KwaZulu-Natal (Figure 5.2); each region comprising habitats differing in their proneness to invasion (Macdonald and Jarman, 1985). The four selected invader species, namely Acacia mearnsii, Lantana camara, Parthenium hysterophorus, and Rubus cuneifolius were noted, from field observations (extensive groundtruthing), to be confined to areas within each of these bioclimatic regions. Given the expanse of each bioclimatic region and the propensity of IAP species to invade areas beyond their current range (Rouget et al., 2004b) each species was modelled across the entire landscape of KwaZulu-Natal.
Figure 5.2: The bioclimatic regions of KwaZulu-Natal (after Phillips, 1973) indicating point locations identifying species‘
92
presence. Insert (a) Acacia mearnsii (KZN midlands), insert (b) Parthenium hysterophorus (northern KZN), insert (c) Rubus cuneifolius (KZN highlands), and insert (d) Lantana camara (coastal plains of KZN).
5.2.1.1 Acacia mearnsii A. mearnsii is a significant invader of ruderal/disturbed areas including grassland, riparian zones, forest gaps, and roadsides throughout its range (warm temperate dry to moist tropical climates) (Duke, 1983; Henderson, 2001). The species‘ favours mesic habitats (altitudinal range of 600-1700m), growing under varied climatic conditions. A mearnsii exhibits a high tolerance to a wide range in precipitation (mean annual precipitation of 660-2280mm), temperature (mean annual temperature of 14.7-27.8°C), and pH (5.0-7.2) (Duke, 1983).
5.2.1.2 Lantana camara L. camara exhibits wide ecological tolerances, evidenced by its diverse geographic distribution and habitat (Day et al., 2003), including, but not limited to, agricultural plots, coastal plains, forest margins, and wetlands (Henderson, 2001).
Preferable habitats
include open, unshaded areas, although disturbed areas such as roadsides, railway tracks, and canals may also prove favourable (Sharma et al., 2005).
L. camara persists in areas characterised by high temperature and rainfall (soils must be well-drained), but rarely proliferates at temperatures below 5°C. In South Africa, the species confines itself to areas maintaining a mean annual surface temperature above 12.5°C (Day et al., 2003).
5.2.1.3 Parthenium hysterophorus Details with regard to the geographical range and distribution of P. hysterophorus, particularly in South Africa, are largely fragmented. The literature provides only sketchy details with reference to the species‘ habitat description and geographical range. This may in part be attributed to P. hysterophorus being an emergent IAP species in South Africa.
The literature that is available indicate that P. hysterophorus grows up to 1.5m (2m in deep rich soils), and proliferates on alkaline, clay-loam to heavy soils, with annual rainfall exceeding 500mm, predominantly in summer (DNRME, 2003; GISD, 2005). P. 93
hysterophorus exhibits highly prolific sexual reproduction, with individuals producing between fifteen thousand and one hundred thousand seeds which may persist in the soil for up to two years (GISD, 2005).
5.2.1.4 Rubus cuneifolius From the literature surveyed, no details were available regarding the geographical range and distribution, habitat description, or ecology of R. cuneifolius in South Africa.
5.2.2 Data acquisition Environmental modelling of species distribution is most often undertaken using climatic variables, with Rouget et al. (2004:476) recognising ‗that the relative importance of climatic variables (is) species-specific‘ and thus ‗generic‘ variables cannot be applied to all species. In this study, selection of climatic variables and physical environmental variables were selected based on their direct and/or indirect influence on species‘ growth and distribution.
Physical data for KZN was acquired as an ESRI shapefile from the KwaZulu-Natal Department of Agriculture and Environmental Affairs, Cedara. The relevant data layers were extracted and exported as raster files using the spatial analyst tools in ArcMap (ESRI, 2008b). Soil depth was selected as an input model parameter given its importance in seed germination and plant growth (Benjamin et al., 2003). Soil depth values were obtained by calculating the average of values from topsoil and subsoil grids. Two topographic variables were included in the study, (1) a 1 minute by 1 minute (1 arc minute ~ 1.7km) digital elevation model (DEM), given its correlation with distribution patterns of species (Torres-Meza et al., 2009); sourced from the South African Atlas of Climatology and Agrohydrology (Schulze et al., 2008), and (2) an aspect grid, used as a proxy to insolation; derived from the DEM using Spatial Analyst tools (ESRI, 2008b).
Climatic data was obtained from the South African Atlas of Climatology and Agrohydrology (Schulze et al., 2008); Schulze et al. (2007) provides details of the methodology employed in deriving the climatic maps. The climate maps comprised one 94
minute by one minute rasters, with rainfall and temperature data having been collected from more than one thousand recording stations across the country over a fifty year period (1950-1999).
Schulze and Maharaj (2007a: 1) affirms that mean annual
temperature ‗represents the very broadest of indices of the environmental status of a location‘, and was thus included as a variable in the analysis given its significance ‗as a general first guide to determine the suitability of a location‘.
Mean annual precipitation, minimum winter temperature, and maximum summer temperature were input to the model, given their known important effects on plant distribution (Woodward, 1987), and their effects on plant respiration, photosynthesis, and growth and development (Schulze and Maharaj, 2007a; Schulze and Maharaj, 2007b). Minimum winter and maximum summer temperature maps were calculated from the monthly means of daily minimum temperature and daily maximum temperature, respectively. For the purpose of this research, winter has been defined as the months of June, July, and August, while summer has been defined as the months of December, January, and February. Inter-annual coefficient of variation of precipitation (CV) was input to the analysis as a ‗best-case‘ scenario of monthly/seasonal fluctuations (Schulze, 2007).
All thematic layers of the explanatory variables (n=8) were projected to WGS84, UTM36S (Hartebeesthoek94 Datum) and resampled (where necessary) to a spatial resolution of 1.7 km, using Spatial Analyst tools (ESRI, 2008a; 2008b). Details of all parameters are provided in Table 5.1.
95
Table 5.1: Input parameters for ecological niche modelling. Biophysical (thematic, categorical) Climatic (continuous, numerical)
Layer Altitude (m) Aspect (degree)
Notes Range: 0 – 3260 Range: -1 – 359.94
Soil depth (cm)
Range: 0 – 51
Mean annual temperature (°C) Mean annual precipitation (mm) Minimum winter temperature (°C)
Range: 7.90 – 22.90 Range: 255.00 – 1923.00 Range: -5.83 – 14.10
Maximum summer temperature (°C) Inter-annual coefficient of variation (%)
Range: 18.63 – 31.97 Range: 18.80 – 42.00
5.2.3 Data processing and analysis 5.2.3.1 Mahalanobis distances Mahalanobis distances provide a powerful tool for quantifying the similarity between an arbitrary set of environmental conditions and an ideal set of conditions (‗optimum‘ conditions that define the species‘ niche) (Farber and Kadmon, 2003; Tsoar et al., 2007). The algorithm incorporates the mean, the variance, as well as the covariance in defining the ellipse, ellipsoid, or hyperellipsoid describing each of the explanatory variables in dimensional space (Jenness, 2003a). The Mahalanobis distance (MD) is the standardised difference between a vector of environmental variables at a point and the mean vector describing all points of species‘ use (Browning et al., 2005; Clark et al., 1993). Application of MD assumes optimal species distribution across available habitats throughout the landscape; habitats being described by a multivariate mean and variance (Knick and Rotenberry, 1998).
Mahalanobis distances were calculated for A. mearnsii, L. camara, P, hysterophorus, and R. cuneifolius using the eight predictor variables described in Table 5.1, and species presence data collected in the field. Mahalanobis distances (between a vector x and a set S of vectors) were calculated as:
D2
x m TC
1
x m 96
Where: D2 = the Mahalanobis distance x = vector indicating (environmental) conditions of a particular site m = vector representing ‗optimum‘ conditions
C
1
= the inverse covariance matrix of independent variables
T = transpose operator S (set of vectors defining habitat suitability) represents environmental conditions for species presence recordings (Farber and Kadmon, 2003). 5.2.3.2 Chi square P-values Assuming a normal distribution of the predictor variables (n), the Mahalanobis distances approximate a Chi-squared distribution with n-1 degrees of freedom (Jenness, 2003a). Pvalues (range: 0-1) were derived from the Mahalanobis distances using the Kappa analysis tools (Jenness, 2003a; 2003b), and subsequently interpreted as analogous to a posterior probability (Rotenberry et al., 2006). P-values allowed for easier interpretation of the maps, as opposed to MD values which can range from almost zero to infinity (Rotenberry et al., 2006).
Cells containing very low p-values (reflecting high MD
values) indicate a high degree of dissimilarity to the ideal value (‗optimum conditions‘ defined by a combination of the input parameters), while p-values close to 1 (reflecting low MD values) indicate a cell is very close to the ideal value.
5.2.4 Model assessment and validation Model accuracy was assessed using the error matrix and reporting on overall accuracy, Kappa, and model specificity (conditional probability of correctly identifying negatives (not suitable sites)). The error matrix was evaluated using 100 random reference points generated in ArcMap (ESRI, 2008b) using Hawths Analysis Tools (Beyer, 2006), and appended to the groundtruth presence/absence point locations collected in the field. Model validation was carried out using the area under the Receiver Operating Characteristic (ROC) curve (AUC) which has been extensively used to validate species‘ distribution models (Elith et al., 2006). The ROC plot relates true positive values (Y97
axis) to false positive values for all thresholds (X-axis) providing a measure of the overall accuracy (Fielding and Bell, 1997), i.e. it compares the predicted suitability map to a reference (presence-absence location) map (Eastman, 2009b). The AUC was calculated as:
Where: xi
= rate of false positives for threshold i
yi
= rate of true positives for threshold i
n+1
= number of thresholds
AUC values range from 0 to 1; a value of 1 indicating perfect discrimination, a value of 0.5 implying predictive discrimination that is no better than a random guess, and values <0.5 indicating performance ‗worse than random‘; a model that fits the data but predicts poorly (Elith et al., 2006).
5.3
Results and Discussion
A visual analysis of the environmental suitability maps show that like L. camara (Figure 5.3: map B), A. mearnsii (Figure 5.3: map A) and R. cuneifolius (Figure 5.3: map D) appear to exhibit a high degree of homeostatic fitness and phenotypic plasticity (Sharma et al., 2005). This may explain their predicted range outside of the bioclimatic region (Phillips, 1973) in which they were sited, and further define the species potential niche (Farber and Kadmon, 2003).
Conversely, the environmental suitability map for P. hysterophorus (Figure 5.3: map C) shows the species to be confined to a locality within the bioclimatic region (lowland woody veld) in which it was sited. This limited range may be explained by the fact that P. hysterophorus is an emergent species in KwaZulu-Natal (and in South Africa) and is still establishing its niche (Foxcroft et al., 2009; Rouget et al., 2004).
98
Overall accuracy ranged from 0.78 (Kappa: 0.67) for P. hysterophorus to 0.93 (Kappa: 0.75) for A. mearnsii (Table 5.2) indicating that the Mahalanobian model has been robust in providing environmental suitability maps delineating predicted habitats based on species‘ presence records and 8 environmental parameters. The high specificity values further confirm power of the MD in defining the potential niche of the species from occurrence data.
99
Figure 5.3: Habitat suitability using Mahalanobian model. Map (A) Acacia mearnsii, map (B) Lantana camara, map (C) Parthenium hysterphorus, and map (D) Rubus cuneifolius. Darker shading of pixels indicates increased habitat suitability.
100
Table 5.2: Model assessment and validation results.
A. mearnsii L. camara P. hysterophorus R. cuneifolius
Accuracy statistics Overall accuracy Kappa Specificity 0.93 0.75 0.83 0.83 0.68 0.74 0.78 0.67 0.65 0.82 0.71 0.71
AUC 0.81 0.65 0.74 0.73
The relative importance of the environmental parameters defining habitat suitability are assumed to be equal for all parameters and for all species modelled. This may not be ideal as a generic set of parameters cannot be readily defined and applied to modelling all species (Rouget et al., 2004). In this study, the relative weights of parameters could not be readily determined and parameter selection was thus driven by the available literature. Clearly, there is need to establish speciesspecific parameters defining the species‘ potential niche as a potential infinite number of combinations of environmental parameters may limit a species‘ range and distribution. Nonetheless, current approaches are useful in providing a general picture of species invasion potential across landscapes (Rouget et al., 2004). In this study, Mahalanobis distances were calculated on 1‘x1‘ grids. Mapping IAP species is always desirable at a fine resolution as coarser resolution impacts on the ability to predict distribution and spread, and to understand and make inferences on the underlying ecology driving the invasion process (Foxcroft et al., 2009). This highlights the need for high resolution, fine scale data.
5.4
Conclusion
An important assumption of the study, as highlighted by Rouget et al. (2004b), is that the potential ranges are governed by the current distribution patterns (from field observations), notwithstanding an over-/under-estimation of species distribution attributed to human-induced disturbance. This assumption is significant with regard to the final output maps from the modelling exercise, particularly that the potential ranges are defined from presence data only, collected during field campaigns, and not accounting for externalities that may impact on the distributions pre and post field observations. Most species‘ distribution models are based on defining climatic variables (Guisan et al., 2006) and fail to account for ecological interactions between the species being modelled and between the
predictor variables of the model (Auranbout et al., 2009; Guisan et al., 2006). While this study incorporated both climatic and physical environmental parameters, Rouget et al. (2004b) suggests the inclusion of landuse change and habitat fragmentation in predictive models of IAP species distribution given their possible effect on dispersal.
Tsoar et al. (2007) further suggest the
incorporation of disturbance, dispersal, and species interaction, as these parameters are often ignored. This warrants investigated through further research.
The overall results of the study, however, demonstrate that ecological niche modelling, based on climatic and biophysical parameters, within a GIS environment, is an effective approach to predicting future invaded environments, and subsequently determining target areas for clearing campaigns. This ultimately forms part of an effective monitoring and control programme for IAP species management. Furthermore, the approach adopted in this study, provides for a relatively simple modelling environment, utilising crude datasets, which could be used generically. The objective of this study was to model the species‘ niche, based upon defining an optimum set of conditions delineated from the mean vector of a combination of environmental parameters based on their Mahalanobis distances. Several authors (e.g. Browning et al., 2005; Calenge et al., 2008; Rotenberry et al., 2006) have modelled the species niche based on defining a minimum set of basic habitat requirements as opposed to the optimum, using partitioned Mahalanobis D2(k).
This
alternative approach needs to be explored further, given that Mahalanobis distances may provide biased predictions of species‘ occurrence under different environmental conditions (Calenge et al., 2008).
102
Chapter 6: General Discussion, Conclusion and Recommendations 6.1
General Discussion and conclusions
This study undertook to identify and map A. mearnsii, L. camara, P. hysterophorus, and R. cuneifolius in four sites across KwaZulu-Natal using medium resolution 10m multispectral SPOT 5 and 15m multispectral ASTER satellite imagery. Several classification decision rules were adopted to mapping the four target species. Although good results were achieved, for example an overall accuracy of 0.8 (Kappa: 0.6) mapping A. mearnsii in the KwaZulu-Natal midlands using SPOT 5 and the ML decision rule, this study highlighted the difficulty in identifying and mapping IAP species using multispectral data in a heterogeneous landscape (Dewey et al., 1991).
Detection and mapping of P. hysterophorus and R. cuneifolius was complicated by the species‘ phenological traits, i.e. winter die-off. Field campaigns established phenological changes for P. hysterophorus and R. cuneifolius during the start of winter, i.e. the two species were noted to exhibit winter die-off. Winter die-off of both P. hysterophorus and R cuneifolius presents a major challenge to its effective detection and mapping using SPOT 5, as this is the time-frame during which SPOT 5 data was collected.
A fundamental principle in remote sensing is that the object being imaged is larger than the smallest mapping unit (Campbell, 2002; Lillesand et al., 2008; Liu and Mason, 2009). In the heterogeneous landscape of KwaZulu-Natal, this certainly is not always the case, nor is there a clearly defined boundary between object classes, i.e. every pixel does not necessarily contain only the material (IAP target species) of interest. The use of fuzzy classifiers could become more mainstream for mapping classes in highly fragmented landscapes.
This research has successfully evaluated the utility of SPOT 5 multispectral data to mapping of selected IAP species, using both statistical and machine learning approaches.
The research
concludes that successful mapping of IAP species is largely dependent on (1) the spatial and spectral resolution of the sensor, (2) the species; patch size, (3) the influence of spectrally-similar materials such as grass, (4) the time of image acquisition, and (5) an understanding of the phenology, ecology, and biology of the target species.
The algorithm and protocol developed for mapping the selected IAP species was designed such that, procedures necessary for the pre-processing of data such as atmospheric correction, geometric 103
correction steps must be performed prior to using the algorithm. The collection of training data that is specific to the scene being processed is crucial.
This study has successfully implemented the Mahalanobian model to predicting habitat suitability (defining species‘ niche habitat) for A. mearnsii, L. camara, P. hysterophorus, and R. cuneifolius in four bioclimatic regions in KwaZulu-Natal, South Africa. The Mahalanobis distance, is a powerful multivariate statistical technique that can be used to predict potential invasion of IAP species by defining the species‘ niche habitat. This has implications for IAP species management, lending itself to strategic monitoring and evaluation planning, and clearing campaigns. This study further highlighted the simple, yet powerful implementation of the Mahalanobis distance statistic by Jenness (2003), within a GIS environment.
6.2
Recommendations
The application of SPOT 5 and ASTER multispectral data to species-level mapping is not without its limitations; primarily spatial and spectral resolution of the sensor. This research has however shown that IAP species can be mapped within reasonable degree of accuracy,where products will subsequently inform the management of IAP specie. There is certainly a wider base for testing its utility on other IAP species, and in other landscapes. Further research on the application of multispectral data to IAP species mapping is certainly worth the initial investment in time and money. This is warranted from an operational directive, particularly with SPOT 5 data being readily available for research and application in higher education and government organisations and facilities.
Furthermore, the availability of a 2.5m pan-sharpened SPOT 5 product presents an enhanced opportunity for further research into detection and mapping of IAP species. The higher spatial resolution should alleviate the influence of mixed pixels, thereby providing for enhanced vegetation discrimination and more accurate mapping. The utility of EO-1 Hyperion hyperspectral data also presents an opportunity for research, although the current high cost of acquisition coupled with resource-consuming processing, makes its general application impractical, particularly for broadscale studies. The recently launched South African satellite (SumbandilaSat) also holds promise for mapping IAP species once its unique data becomes available. Data fusion of high spatial resolution multispectral data with high spectral resolution (hyperspectral) data may be the ideal data source for IAP identification, discrimination, and mapping. 104
Results from validating the developed mapping algorithm indicate that the mapping algorithm is best suited for semi-automatic mapping procedure since training data specific to the image data being processed is required prior to using. The efficacy of using e-Cognition and syntactic approaches to classification vis-à-vis ArcGIS-ArcEditor and statistics-based Maximum Likelihood classification to develop protocols and algorithms for mapping IAP species is recommended for investigation. Furthermore, a combination of Rule-based and syntactic approaches to classification offers the potential to develop classification protocols and algorithms, which may offer a possibility to off-set the need for generating training data. Thus without the need to generate training data, algorithms developed from such classification protocols may be fully automated.
Further research is certainly required to increase our knowledge with regards to IAP species‘ ecology and biology. This information will facilitate a better understanding into species‘ invasion, spread, range, and distribution. Furthermore, increased data collection, housing, management, and sharing, is more than likely to benefit both current and future research.
Future research priorities should further address the development of a predictive understanding of rates of spread of IAP species (Richardson and van Wilgen, 2004). Modelling the rate of spread of an invasive plant population, coupled with predictive models of habitat suitability (species‘ niche) would provide an effective mechanism for the early detection, rapid response, and eradication of IAP species (Lodge et al., 2006), as part of an integrated IAP species management programme.
Research opportunities exist with the availability of annual acquisitions of SPOT 5 imagery, including change detection and time-series analysis to determine rate of spread, gain a better understanding of spread dynamics, evaluate monitoring and clearing programmes, and strategise IAP species management planning. The utility of SPOT 5 data, GPS, and GIS and RS techniques would thus form an integral component of integrated IAP species management.
With the number of new emergent species on the increase, shortfalls in funding, and imminent threats posed by IAP species, South Africa needs to adopt and impose clear and stringent directives as to the management and control, eradication, and prevention against entry into the country of these non-native species. Given the inextricable link between humans and the natural environment, the pursuit of human endeavour will always be associated with habitat alteration, landscape fragmentation, climate change, and subsequently, the need for IAP species management.
105
106
References: Abdi, H., 2003. Neural networks. In Lewis-Beck, M., A. Bryman, T. Fluting (Eds.), ‗The Sage encyclopaedia of social sciences research methods‘. Thousand Oaks: Sage. Abraham, A., 2005. Artificial neural networks. In Sydenham, P.H., R. Thorn (Eds.), ‗Handbook of measuring system design‘. Chichester: John Wiley and Sons. Aitkenhead, M.J., and I.H. Aalders, 2008. Classification of Landsat thematic imagery for land cover using neural networks. International Journal of Remote Sensing 29 (7), 2075-2084. Albañez-Lucero, M.O., and F. Arreguín-Sánchez, 2009. Modelling the spatial distribution of red grouper (Epinephelus morio) at Campeche Bank, México, with respect substrate. Ecological Modelling 220, 2744-2750. Anderson, G.L., J.H. Everitt, D.E. Escobar, N.R. Spencer, and R. Andrascik, 1996. Mapping leafy spurge (Euphorbia esula) infestations using aerial photography and geographic information systems. Geocarto International 11, 81-89. Anderson, R.P., D. Lew, and A.T. Peterson, 2003. Evaluating predictive models of species‘ distributions: criteris for selecting optimal models. Ecological Modelling 162, 211-232. Anderson, R.P., and E. Martinez-Meyer, 2004. Modelling species‘ geographic distributions for preliminary conservation assessments: an implementation with the spiny pocket mice (Heteromys) of Ecuador. Biological Conservation 116, 167-179. Atkinson, P.M., M.E.J. Cutler, and H.G. Lewis, 1997. Mapping sub-pixel proportional land cover with AVHRR imagery. International Journal of Remote Sensing 18, 917-935. Atkinson, P.M., and A.R.L. Tatnall, 1997.
Introduction: Neural networks in remote sensing.
International Journal of Remote Sensing 18 (4), 699-709. Aurambout, J.P., K.J. Finlay, J. Luck, and G.A.C. Beattie, 2009. A concept model to estimate the potential distribution of the Asiatic citrus psyllid (Diaphorina citri Kuwayama) in Australia under climate change – A means for assessing biosecurity risk. Ecological Modelling 220, 2512-2524. Barnett, D.T., T.J. Stohlgren, C.S. Jarnevich, G.W. Chong, J.A. Ericson, T.R. Davern, and S.A. Simonson, 2007. The art and science of weed mapping.
Environmental Monitoring and
Assessment 132, 235-252. Basheer, I.A., and M. Hajmeer, 2000. Artificial neural networks: fundamentals, computing, deign, and application. Journal of Microbiological Methods 43, 3-31. Beaumont, L.J., L. Hughes, and M. Poulsen, 2005. Predicting species distributions: use of climatic parameters in BIOCLIM and its impact on predictions of species‘ current and future 107
distributions. Ecological Modelling 186, 250-269. Benediktsson, J.A., P.H. Swain, and O.K. Esroy, 1993. Conjugate-gradient neural networks in classification of multisource and very-high-dimensional remote sensing data. International Journal of Remote Sensing 14 (15), 2883-2903. Benjamin, J.G., D.C. Nielsen, and M.F. Vigil, 2003. Quantifying effects of soil conditions on plant growth and crop production. Geoderma 116, 137-148. Beyer, H.L., 2006. Hawth's Analysis Tools for ArcGIS. Available from: http://www.spatialecology.com/htools [Accessed 20 January 2007] Boylen, C.W., L.W. Eichler, J.S. Bartkowski, and S.M. Shaver, 2006.
Use of geographic
information systems to monitor and predict non-native aquatic plant dispersal through northeastern North America. Hydrobiologia 570, 243-248. Bradley, B.A., and J.F. Mustard, 2006. Characterising the landscape dynamics of an invasive plant and risk of invasion using remote sensing. Ecological Applications 16 (3), 1132-1147. Bredenkamp, G.J., J.E. Granger, R. Lubke, and N. van Rooyen, 1998. Moist upland grassland. In Low, A.B., A.G. Rebelo (Eds.), ‗Vegetation of South Africa, Lesotho and Swaziland‘, 2nd Ed. Department of Environmental Affairs and Tourism, Pretoria. Bridges, D.C., 1994. Impact of Weeds on Human Endeavours. Weed Technology 8 (10), 392-395. Bromilow, C., 2001. Problem plants of South Africa: A guide to the identification and control of more than 300 invasive plants and other weeds. Pretoria: Briza Publications. Brooks, M.L., C.M. D‘Antonio, D.M. Richardson, J.B. Grace, J.E. Keeley, J.M. DiTomaso, R.J. Hobbs, M. Pellant, and D. Pyke, 2004.
Effects of invasive alien plants on fire regimes.
BioScience 54 (7), 677-688. Browning, D.M., S.T. Beaupre, and L. Duncan, 2005. Using partitioned Mahalanobis D2(k) to formulate a GIS-based model of timber rattlesnake hibernacula.
The Journal of Wildlife
Management 69 (1), 33-44. Calenge, C., G. Darmon, M. Basille, A. Loison, and J.-M. Jullien, 2008.
The factorial
decomposition of the Mahalanobis distances in habitat selection studies. Ecology 89 (2), 555566. Camp, K.G., 1997. The bioresource groups of KwaZulu-Natal. Natural Resource Section, Cedara. Cedara report No. N/A/97/6. Campbell, J.B., 2002. Introduction to Remote Sensing, 3RD Ed. New York: Guilford Press. CARA, 1983. Conservation of Agricultural Resources Act, Act 43 of 1983. Pretoria, South Africa. Chamberlain, D., H. Essop, C. Hougard, S. Malherbe, and R. Walker, 2005. Genesis Report: Part 1. The contribution, costs and development opportunities of the forestry, timber , pulp and paper industries in South Africa. Genesis Analytics (Pty) Ltd, Johannesburg, South Africa. 108
Chang, K-T., 2010. Introduction to geographic information systems, 5TH Ed. New York: McGrawHill. Chapman, R.A., and D.C. Le Maitre 2001. Scenarios for Alien Invading Woody Plants. Water Research Commission. Research Report No. 907/1/01, Pretoria, South Africa. Clark, J.D., J.E. Dunn, and K.G. Smith, 1993. A multivariate model of female black bear habitat use for a geographic information system. Journal of Wildlife Management 57, 519-526. Congalton, R.G., and K. Green, 2009. Assessing the accuracy of remotely sensed data: principles and practices. Boca Raton, Florida: Taylor & Francis. Cooper, M.R., and M.D. Cooper, 2002. The Emperor moths of KwaZulu-Natal. Pietermaritzburg: Peroniceras Press. Cuddington, K., and A. Hastings, 2004. Invasive engineers. Ecological Modelling 178, 335-347. Cullis, J.D.S., A.H.M. Görgens, and C. Marias, 2007. A strategic study of the impact of invasive alien plants in the high rainfall catchments and riparian zones of South Africa on total surface water yield. Water SA 33 (1), 35-42. Davis, M.A., J.P. Grime, and K. Thompson, 2000. Fluctuating resources in plant communities: A general theory of invisibility. Journal of Ecology 88, 528-534. Day, M.D., C.J. Wiley, J. Playford, and M.P. Zalucki, 2003. Lantana: Current management status and future prospects. Australian Centre for International Agricultural Research, Canberra, Australia. De Neergaard, A., C. Saarnak, T. Hill, M. Khanyile, A.M. Berzosa, and T. Birch-Thomsen, 2005. Australian wattle species in the Drakensberg region of South Africa – An invasive alien or a natural resource? Agricultural Systems 85, 216-233. DEAT, 2009. List of prohibited alien species in terms of Section 67: NEMBA, 2004. Department of Environmental Affairs and Tourism, Government Gazette No. 32090. Dewey, S.A., K.P. Price, and D. Ramsey, 1991. Satellite remote sensing to predict potential distribution of dyers wood (Isatis tinctoria). Weed Technology 5, 479-484. Dixon, B, and N. Candade, 2008. Multispectral landuse classification using neural networks and support vector machines: one or the other, or both? International Journal of Remote Sensing 29 (4), 1185-1206. DNRME, 2003.
Parthenium weed management: Challenges, opportunities, and strategies.
Department of Natural Resources, Mines and Energy: The State of Queensland. Draper, D., A. Rosselló-Graell, C. Garcia, C.T. Gomes, and C. Sérgio, 2003. Application of GIS in plant conservation programmes in Portugal. Biological Conservation 113, 337-349. Duke, J.A., 1983. Handbook of energy crops. Unpublished. Available
from:
http://www.hort.purdue.edu/newcrop/duke_energy/Acacia_mearnsii.html 109
[Accessed 31 August 2009] Dye, P., and C. Jarmain, 2004. Water use by black wattle (Acacia mearnsii): implications for the link between removal of invading trees and catchment streamflow response. South African Journal of Science 100, 40-44. Eastman, J.R., 2009a. IDRISI 16.03: The Taiga Edition. Worcester, Massachesettes: Clark Labs, Clark University. Eastman, J.R., 2009b.
IDRISI Taiga guide to GIS and image processing.
Worcester,
Massechusetts: Clark Labs, Clark University. Eastman, J.R., 2009c. MLP. Accessed in IDRISI Taiga on-line help system, IDRISI: The Taiga Edition. Worcetser, Massachusetts: Clark Labs, Clark University. Elith, J., C.H. Graham, R.P. Anderson, M. Dudík, S. Ferrier, A. Guisan, R.J. Hijmans, F. Huettmann, J.R. Leathwick, A. Lehmann, J. Li, L.G. Lohmann, B.A. Loiselle, G. Manion, G. Moritz, M. Nakamura, Y. Nakazawa, J.McC. Overton, A. T. Peterson, S.J. Phillips, K. Richardson, R. Scachetti-Pereira, R.E. Schapire, J. Soberón, S. Williams, M.S. Wisz, and N.E. Zimmermann, 2006.
Novel methods improve prediction of species‘ distributions from
occurrence data. Ecography 29, 129-151. Enright, W.D., 2000. The Effect of Terrestrial Invasive Alien Plants on Water Scarcity in South Africa. Physics and Chemistry of the Earth, Part B: Hydrology, Oceans and Atmosphere 25 (3), 237-242. ENVI, 2006. ENVI Help. Research Systems, Inc., Boulder, CO, USA. ERDAS, 2008a. ERDAS field guide, vol.2. ERDAS, Inc., Norcross, Georgia, USA. ERDAS, 2008b. ERDAS IMAGINE. ERDAS, Inc., Norcross, Georgia, USA. ERDAS, 2008c. IMAGINE Subpixel Classifier: User‘s guide. ERDAS, Inc., Norcross, Georgia, USA. ESRI, 2002. ArcView GIS 3.3. Environmental Systems Research Institute Inc. (ESRI), Redlands, California, USA. ESRI, 2008a. ArcGIS Desktop Help. Environmental Systems Research Institute Inc. (ESRI), Redlands, California, USA. ESRI, 2008b. ArcMap 9.3 (ArcEditor). Environmental Systems Research Institute Inc. (ESRI), Redlands, California, USA. Ewel, JJ, DJ O‘Dowd, J Bergelson, CC Daehler, CM D‘Antonio, LD Gómez, DR Gordon, RJ Hobbs, A Holt, KR Hopper, CE Hughs, M LaHart, RRB Leakey, WG Lee, LL Loope, DH Lorence, SM Louda, AE Lugo, PB McEvoy, DM Richardson and PM Vitousek 1999. Deliberate Introductions of Species: Research Needs. BioScience 49 (8), 619-630. Fairbanks, D.H.K., and G.A. Benn, 2000.
Identifying regional landscapes for conservation 110
planning: a case study from KwaZulu-Natal, South Africa. Landscape and Urban Planning 50, 237-257. Farber, O., and R. Kadmon, 2003. Assessment of alternate approaches for bioclimatic modelling with special emphasis on the Mahalanobis distance. Ecological Modelling 160, 115-130. Fielding, A.H., and J.F. Bell, 1997. A review of methods for the assessment of prediction errors in conservation presence/absence models. Environmental Conservation 24, 38-49. Foody, G.M., 1995. Land cover classification using an artificial neural network with ancillary information. International Journal of Geographical Information Systems 9 (5), 527-542. Foody, G.M., 2006. Sub-pixel methods in remote sensing. In de Jong, S.M. and F.D. van der Meer (Eds.) ‗Remote Sensing Image Analysis: Including the Spatial Domain‘. Netherlands: Springer. Foody, G.M., 2001. Thematic mapping from remotely sensed data with neural networks: MLP, RBF and PNN based approaches. Journal of Geographical Systems 3, 217-232. Foody, G.M., D.S. Boyd, and C. Sanchez-Hernandez, 2007. Mapping a specific class with an ensemble of classifiers. International Journal of Remote Sensing 28 (8), 1733-1746. Foody, G.M., A. Mathur, C. Sanchez-Hernandez, and D.S. Boyd, 2006.
Training set size
requirements for the classification of a specific class. Remote Sensing of Environment 104, 114. Foxcroft, L.C., D.M. Richardson, M. Rouget, and S. MacFayden, 2009. Patterns of alien plant distribution at multiple spatial scales in a large national park: implications for ecology, management and monitoring. Diversity and Distributions 15, 367-378. Frenot, Y., J.C. Gloaguen, L. Massé, and M. Lebouvier 2001. Human Activities, Ecosystem Disturbance and Plant Invasions in Subantarctic Crozet, Kerguelen and Amsterdam Islands. Biological Conservation 101:33-50. Gardner, M.W., and S.R. Dorling, 1998. Artificial neural networks (the multilayer perceptron) – A review of applications in the atmospheric sciences. Atmospheric Environment 32 (14/15), 26272636. Gillham, J.H., A.L. Hild, J.H. Johnson, E.R. Hunt Jr., and T.D. Whitson, 2004. Weed invasion susceptibility prediction (WISP) model for use with geographic information systems. Arid Land Research and Management 18, 1-12. Giovanelli, J.G.R., M.F. de Siqeira, C.F.B. Haddad, and J. Alexandrino, 2010.
Modelling a
spatially restricted distribution in the Neotropics: How the size of calibration area affects tyhe performance of five presence-only methods. Ecological Modelling 221, 215-224. GISD (Global Invasive Species Database), 2005. Parthenium hysterophorus. Available from: http://www.issg.org/database/species/ecology.asp?si=153andfr=1andsts=sssandlang=EN 111
[Accessed 22 May 2009]. Govender, A., 2007. Personal communication. Clearing team‘s area manager for Greytown: Umgeni Water. Granger, J.E., and G.J. Bredenkamp, 1998. Short mistbelt grassland. In Low, A.B., A.G. Rebelo (Eds.), ‗Vegetation of South Africa, Lesotho and Swaziland‘, 2nd Ed.
Department of
Environmental Affairs and Tourism, Pretoria. Granger, J.E., G.J. Bredenkamp, and N. van Rooyen, 1998. Coastal bushveld-grassland. In Low, A.B., A.G. Rebelo (Eds.), ‗Vegetation of South Africa, Lesotho and Swaziland‘, 2nd Ed. Department of Environmental Affairs and Tourism, Pretoria. Guisan, A., A. Lehman, S. Ferrier, M. Austin, J.MC.C. Overton, R. Aspinall, and T. Hastie, 2006. Making better biogeographical predictions of species‘ distributions. Journal of Applied Ecology 43, 386-392. Hellgren, E.C., S.L. Bales, M.S. Gregory, D.M. Leslie Jr., and J.D. Clark, 2007.
Testing a
Mahalanobis distance model of black bear habitat use in the Ouachita Mountains of Oklahoma. The Journal of Wildlife Management 71 (3), 924-928. Henderson, L., 1998. Southern African Plant Invaders Atlas (SAPIA). Applied Plant Science 12, 31-32. Henderson, L., 1999. The Southern African Plant Invaders Atlas (SAPIA) and its contribution t biological weed control. African Entomology Memoir 1, 159-163. Henderson, L., 2001. Alien Weeds and Invasive Plants: A Complete Guide to Declared Weeds and Invaders in South Africa. Plant Protection Research Institute, Agricultural Research Council. Hengl, T., H. Sierdsema, A. Radović, and A. Dilo, 2009. Spatial prediction of species‘ distributions from occurrence-only records: combining point pattern analysis, ENFA, and regression-kriging. Ecological Modelling 220, 3499-3511. Hirzel, A.H., J. Hausser, D. Chessel, and N. Perrin, 2002. Ecological-niche factor analysis: How to compute habitat-suitability maps without absence data? Ecology 83 (7), 2027-2036. Impson, F.A.C., V.C. Moran, and J.H. Hoffman, 2004. Biological control of an alien tree, Acacia cyclops, in South Africa: impact and dispersal of a seed-feeding weevil, Melanterius servulus. Biological Control 29, 375-381. IUCN, 2000. IUCN Guidelines for the Prevention of Biodiversity Loss caused by Alien Invasive Species. Fifth meeting of the Conference of the Parties to the Convention on Biological Diversity. Nairobi, Kenya 15-26 May 2000. Jain, A.K., J. Mao, and K.M. Mohiuddin, 1996. Artificial neural networks: A tutorial. Computer 29 (3), 31-44. Jenness, J., 2003a. Mahalanobis distances. Arizona: Jenness Enterprises. 112
Jenness, J., 2003b. Mahalanobis distances (mahalanobis.avx) extension for ArcView 3.x. Jenness Enterprises. Available from: http://jennessent.com/arcview/mahalanobis.htm [Accessed 23 July 2007] Jenness, J., and J.J. Wynne, 2007a.
Cohen’s Kappa and classification table metrics 2.1: An
ArcView extension for accuracy assessment of spatially-explicit models. Arizona: Jenness Enterprises. Jenness, J., and J.J. Wynne, 2007b. Kappa analysis (kappa_stats.avx) extension for ArcView 3.x. Jenness Enterprises. Available from: http://jennessent.com/arcview/kappa_stats.htm [Accessed 23 July 2007] Joshi, C., J. de Leeuw, and I.C. van Duren, 2004. Remote Sensing and GIS Applications for Mapping and Spatial Modelling of Invasive Species.
Proceedings of ISPRS 35, B7 (7),
Istanbul. 669-677. Kaiser, J., 1999. Stemming the tide of invading species. Science 285 (5435), 1836-1841. Kanellopoulos, I., and G.G. Wilkinson, 1997. Strategies and best practice for neural network image classification. International Journal of Remote Sensing 18 (4), 711-725. Kavzoglu, T., 2009. Increasing the accuracy of neural network classification using refined training data. Environmental Modelling and Software 24, 850-858. Kotiluoto, R., K. Ruokolainen, and M. Kettunen, 2008. Invasive Acacia auriculiformis Benth. in different habitats in Unguja, Zanzibar. African Journal of Ecology 47, 77-86. Knick, S.T., and J.T. Rotenberry, 1998.
Limitations to mapping habitat use areas in changing
landscapes using the Mahalanobis distance statistic. Journal of Agricultural, Biological, and Environmental Statistics 3 (3), 311-322. Kokaly, R.F., D.G. Despain, R.N. Clark, and K.E. Livo, 2003. Mapping vegetation in Yellowstone National Park using spectral feature analysis of AVIRIS data. Remote Sensing of Environment 84, 437-456. Kolar, C.S., and D.M. Lodge, 2001. Progress in Invasion Biology: Predicting Invaders. Trends in Ecology and Evolution 16 (4), 199–204. Krogh, A., 2008. What are artificial neural networks? Nature Biotechnology 26 (2), 195-197. Lawrence, R.L., S.D. Wood, and R.L. Sheley, 2006. Mapping invasive plants using hyperspectral imagery and Breiman Cutler classifications (RandomForest). Remote Sensing of Environment 100, 356-362. Le Maitre, D.C., B.W. van Wilgen, C.M. Gelderblom, C. Bailey, R.A. Chapman, and J.A. Nel, 2002. Invasive alien trees and water resources in South Africa: Case studies of the costs and benefits of management. Forest Ecology and Management 160, 143-159. Le Maitre, D.C., B.W. van Wilgen, R.A. Chapman, and D.H. McKelly, 1996. Invasive plants and 113
water resources in the Western Cape Province, South Africa: Modelling the consequences of a lack of management. Journal of Applied Ecology 33, 161-172. Le Maitre, D.C., D.B. Versfeld, and R.A. Chapman, 2000. The impact of invading alien plants on surface water resources in South Africa: A preliminary assessment. Water SA 26 (3), 397-408. Lillesand, T.M., R.W. Kiefer, and J.W. Chipman, 2008. Remote sensing and image interpretation, 6th Ed. Hoboken, New Jersey: John Wiley & Sons Inc. Liu J.G. and P.J. Mason, 2009.
Essential image processing and GIS for remote sensing.
Chichester, UK: John Wiley & Sons Ltd. Liu, W., and E.Y. Wu, 2005. Comparison of non-linear mixture models: a sub-pixel classification. Remote Sensing of Environment 94, 145-154. Lodge, D.M., S. Williams, H.J. macIsaac, K.R. Hayes, B. Leung, S. Reichard, R.N. Mack, P.B. Moyle, M. Smith, D.A. Andow, J.T. Carlton, and A. McMichael, 2006. Biological invasions: recommendations for U.S. policy and management. Ecological Applications 16 (6), 2035-2054. Lonsdale, W.M., 1999. Global patterns of plant invasions and the concept of invasibility. Ecology 80, 1522-1536. Los, S.O., C.J. Tucker, A. Anyamba, M. Cherlet, G.J. Collatz, L. Giglio, F.G. Hall, and J.A. Kendall, 2002. Environmental modelling with GIS and Remote Sensing. London: Taylor and Francis. Low, A.B., and A.G. Rebelo (Eds.), 1998. Vegetation of South Africa, Lesotho and Swaziland: A companion to the vegetation map of South Africa, Lesotho and Swaziland. Pretoria: Department of Environmental Affairs and Tourism. Lu, D., and Q. Weng, 2007. A survey of image classification methods and techniques for improved classification performance. International Journal of Remote Sensing 28 (5), 823-870. MacDonald, I.A.W., J.K. Reaser, C. Bright, L.E. Neville, G.W. Howard, S.J. Murphy, and G. Preston (Eds.). 2003. Invasive alien species in southern Africa: national reports and directory of resources. Global Invasive Species Programme, Cape Town, South Africa. Macdonald, I.A.W., and M.L. Jarman (Eds.), 1985.
Invasive alien plants in the terrestrial
ecosystems of Natal, South Africa. South African National Scientific Programmes Report No. 118, Council for Scientific and Industrial Research, Pretoria. Madden, M., 2004. Remote sensing and geographic information system operations for vegetation mapping of invasive exotics. Weed Technology 18, 1457-1463. Marais, C., B.W. van Wilgen, and D. Stevens., 2004. Clearing of invasive alien plants in South Africa : a preliminary assessment of costs of progress. South African Journal of Science 100, 97-103. Mas, J.F., and J.J. Flores, 2008. The application of artificial neural networks to the analysis of 114
remotely sensed data. International Journal of Remote Sensing 29 (3), 617-663. Mau-Crimmins, T.M., and B.J. Orr, 2005.
Monitoring invasive plants using hand-held GIS
technology. USDA Forest Service Proceedings RMRS-P-36, 298-301. Melgani, F., and L. Bruzzone, 2004. Classification of hyperspectral remote sensing images with support vector machines. IEEE Transactions on Geoscience and Remote Sensing 42 (8), 17781790. Miller, D.M., E.J. Kaminsky, and S. Rana, 1995. Neural network classification of remote sensing data. Computers and Geosciences 21, 377-386. Mills, H., M.E.J. Cutler, and D. Fairbairn, 2006. Artificial neural networks for mapping regionalscale upland vegetation from high spatial resolution imagery. International Journal of Remote Sensing 27 (11), 2177-2195. Mooney, H.A., and E.E. Cleland, 2001. The evolutionary impact of invasive species. Procedures of National Academy of Science 98 (10), 5446-5451. Mooney, H.A., and R.J. Hobbs, 2000. Invasive Species in a Changing World. Washington, D.C.: Island Press. Moran, V.C., J.H. Hoffman, D. Donnelly, B.W. van Wilgen, and H.G. Zimmerman, 2000. Biological control of alien, invasive pine trees (Pinus species) in South Africa. In N.R. Spencer (Ed.) ‗Proceedings of the X International Symposium on Biological Control of Weeds‘, 941953. Müllerova, J., P. Pyšek, V. Jarošik, and J. Pergl, 2005. Aerial photographs as a tool for assessing the regional dynamics of the invasive plant species Heracleum mantegazzium. Journal of Applied Ecology 42, 1042-1053. Narumalani, S., D.R. Mishra, R. Wilson, P. Reece, and A. Kohler, 2009. Detecting and mapping our invasive species along the floodplain of North Platte River, Nebraska. Weed Technology 23, 99-107. NAS (National Academy of Sciences), 2002. Predicting invasions of nonindigenous plants and plant pests.
Committee on the scientific basis for predicting the invasive potential of
nonindigenous plants and plant pests in the United States, National Research Council. National Academy Press: Washington D.C. NEMBA, 2004. National Environmental Management Biodiversity Act, Act 10 of 2004. Pretoria, South Africa. Nel, S., 2008. Global positioning: technologies and performance. Hoboken, New Jersey: WileyInterscience. Nel, J.L., D.M. Richardson, M. Rouget, T.N. Mgidi, N. Mdzeke, D.C. Le Maitre, B.W. van Wilgen, L. Schonegevel, L. Henderson, and S. Neser, 2004. A proposed classification of invasive alien 115
plant species in South Africa: towards prioritising species and areas for management action. South African Journal of Science 100, 53-64. Okin, G.S., D.A. Roberts, B. Murray, and W.J. Okin, 2001. Practical limits on hyperspectral vegetation discrimination in arid and semiarid environments. Remote Sensing of Environment 77, 212-225. Pearson, R.G., T.P. Dawson, M.P. Berry, 2002. SPECIES: a spatial evaluation of climate impact on the envelope of species. Ecological Modelling 154, 298-300. Phillips, J., 1973. The agricultural and related development of the Tugela basin and its influent surrounds. Town and Regional Planning report, Vol. 19. Natal Town and Regional Planning commission, Pietermaritzburg. Phillips, S.J., R.P. Anderson, and R.E. Schapire, 2006. Maximum entropy modelling of species geographic distributions. Ecological Modelling 190, 231-259. Pimentel, D., L. Lach, R. Zuniga, and D. Morrison, 2005. Update on the environmental and economic costs associated with alien-invasive species in the United States.
Ecological
Economics 52, 273-288. Pino, J., X. Font, J. Carbó, M. Jové, and L. Pallarès, 2005. Large-scale correlates of alien plant invasion in Catalonia (NE of Spain). Biological Conservation 122, 339-350. Pritekel, C., A. Whittemore-Olson, N. Snow, and J.C. Moore, 2006. Impacts from invasive plant species and their control on the plant community and belowground ecosystem at Rocky Mountain National Park, USA. Applied Soil Ecology 32, 132-141. Qiu, F. and J.R. Jensen, 2004. Opening the black box of neural networks for remote sensing image classification. International Journal of Remote Sensing 25 (9), 1749-1768. Raghubanshi, A.S., L.C. Rai, J.P. Gaur, and J.S. Singh, 2005.
Invasive alien species and
biodiversity in India. Current Science 88 (4), 539-540. Rejmánek, M., 2000. Invasive plants: approaches and predictions. Austral Ecology 25, 497-506. Richards, J.A., 1986. Remote sensing digital image analysis: an introduction. Berlin: Springer. Richardson D.M., and B.M. van Wilgen, 2004. Invasive alien plants in South Africa: how well do we understand the ecological impacts? South African Journal of Science 100, 45-52. Rotenberry, J.T., K.L. Preston, and S.T. Knick, 2006. GIS-based niche modelling for mapping species‘ habitat. Ecology 87 (6), 1458-1464. Rouget, M., D.M. Richardson, S.J. Milton, and D. Polakow, 2004a. Predicting invasion dynamics of four alien Pinus species in a highly fragmented semi-arid shrubland in South Africa. Plant Ecology 152, 79-92. Rouget, M., D.M. Richardson, J. Nel, D.C. Le Maitre, B. Egoh, and T. Mgidi, 2004b. Mapping the potential ranges of major plant invaders in South Africa, Lesotho and Swaziland using climatic 116
suitability. In ‗An assessment of invasion potential of invasive alien plant species in South Africa‘. CSIR Report No. ENV-S-C 2004-108. Rowlinson, L.C., M. Summerton, and F. Ahmed, 1999. Comparison of remote sensing data sources and techniques for identifying and classifying alien invasive vegetation in riparian zones. Water SA 25 (4), 497-500. Sanz-Elorza, M., E.D. Dana, and E. Sobrino, 2006. Invasibility of an inland area in NE Spain by alien plants. Acta Oecologica 29, 114-122. Schalkoff, R., 1992. Pattern Recognition: statistical, structural and neural approaches. New York: John Wiley and Sons, Inc. Schalkoff, R.J., 1997. Artificial neural networks. New York: McGraw-Hill. Schulze, R.E., 2007. Coefficient of variation of annual precipitation. In R.E. Schulze (Ed.) ‗South African Atlas of Climatology and Agrohydrology‘. Water Research Commission, Technical Report No. 1489/1/06, Pretoria. Schulze, R.E., and M. Maharaj, 2007a. Daily minimum temperatures. In R.E. Schulze (Ed) ‗South African atlas of climatology and agrohydrology‘. Water Research Commission, Technical Report No. 1489/1/06, Pretoria. Schulze, R.E., and M. Maharaj, 2007b. Daily maximum temperatures. In R.E. Schulze (Ed) ‗South African atlas of climatology and agrohydrology‘. Water Research Commission, Technical Report No. 1489/1/06, Pretoria. Schulze, R.E., M. Maharaj, M.L. Warburton, C.J. Gers, M.J.C. Horan, R.P. Kunz, and D.J. Clark, 2008. South African atlas of climatology and agrohydrology. Water Research Commission Report, 1489/1/08. Schulze, R.E., M. Maharaj, S.D. Lynch, B.J. Howe, and B. Melvil-Thomson, 1997. South African atlas of agrohydrology and climatology. Water Research Commission Report, TT82/96. Sharma, G.P., A.S. Raghubanshi, and J.S. Singh, 2005. Lantana invasion: An overview. Weed Biology and Management 5, 157-165. Shea, K., and P. Chesson, 2002.
Community ecology theory as a framework for biological
invasions. Trends in Ecology and Evolution 17 (4), 170-176. Sihlangu, S., 2007.
Personl communication.
Clearing team‘s area manager, Richards Bay-
Empangeni, Umhlathuze Municipality. Skidmore, A., 2002. Environmental modelling with GIS and remote sensing: introduction. In Skidmore, A. (Ed.) ‗Environmental Modelling with GIS and Remote Sensing‘. London: Taylor and Francis. Skov, F., and J-C Svenning, 2003. Predicting plant species richness in a managed forest. Forest Ecology and Management 180, 583-593. 117
Soberón, J., 2007. Grinnelian and Eltonian niches and geographic distributions of species. Ecology Letters 10, 1115-1123. Stow, D., Y. Hamada, L. Coulter, and Z. Anguelova, 2009. Monitoring shrubland habitat changes through object-based change identification with airborne multispectral imagery.
Remote
Sensing of Environment 112, 1051-1061. Sutherst, R.W., G.F. Maywald, T. Yonow, and P.M. Stevens, 1999.
CLIMEX user guide –
predicting the effects of climate on plants and animals. Collingwood, Victoria, Australia: CSIRO Publishing. Swain, P.H., and S.M. Davis (Eds.), 1978. Remote Sensing: the quantitative approach. New York: McGraw-Hill. Swain, P.H., and R.C. King, 1973. Two effective feature selection criteria for multispectral remote sensing. LARS technical note 042673. Presented at The International Joint Conference on Pattern Recognition, Washington, D.C. Tassin, J., J.M. Médoc, C.A. Kull, J.N. Rivière, and G. Balent, 2009. Can invasion patches of Acacia mearnsii serve as colonising sites for native plant species on Rèunion (Mascarene archipelago)? African Journal of Ecology 47, 422-432. Thiébaut, G., 2005. Does competition for phosphate supply explain the invasion pattern of Elodea species? Water Research 39, 3385-3393. Thompson, K., J.G. Hodgson, J.P. Grime, and M.J.W. Burke, 2001. Plant traits and temporal scale: Evidence from a 5-year invasion experiment using native species. Journal of Ecology 89, 10541060. Thuiller, W., D.M. Richardson, M. Rouget, S. Procheş, and J.R.U. Wilson, 2006. Interactions between environment, species traits, and human uses describe patterns of plant invasions. Ecology 87 (7), 1755-1769. Torres-Meza, M.de J., Báez-González, L.H. Maciel-Pérez, E. Quezada-Guzmán, and J.S. SierraTristán, 2009. GIS-based modelling of the geographic distribution of Quercus emoryi Torr. (Fagaceae) in México and identification of significant environmental factors influencing the species‘ distribution. Ecological Modelling 229, 3599-3611. Tou, J.T., and R.C. Gonzalez, 1974. Pattern recognition principles. London: Addison-Wesley. Travaini, A., J. Bustamante, A. Rodríguez, S. Zapata, D. Procopio, J. Pedrana, and R.M. Peck, 2007. An integrated framework to map animal distributions in large and remote regions. Diversity and Distributions 13, 289-298. Tsoar, A., O. Allouche, O. Steinitz, D. Rotem, and R. Kadmon, 2007. A comparative evaluation of presence-only methods for modelling species distribution. Diversity and Distributions 13, 397405. 118
Umgeni Water, 1997. Riparian zone rehabilitation study – Midmar Dam Catchment. Umgeni Water Report M0480/1, MBB Consulting Engineers Inc., Pietermaritzburg. Underwood, E., S. Ustin, and D. DiPietro, 2003. Mapping nonnative plants using hyperspectral imagery. Remote Sensing of Environment 86, 150-161. Vaclavik, T., and R.K. Meetenmeyer, 2009. Invasive species distribution modelling (iSDM): Are absence data and dispersal constraints needed to predict actual distributions?
Ecological
modelling 220, 3248-3258. Van Rooyen, N., and G. Bredenkamp, 1998. Lebombo arid mountain bushveld. In Low, A.B., A.G. Rebelo (Eds.), ‗Vegetation of South Africa, Lesotho and Swaziland‘, 2nd Ed. Department of Environmental Affairs and Tourism, Pretoria. Van Wilgen, B.W., 2009. The evolution of fire and invasive alien plant management practices in fynbos. South African Journal of Science 105, 335-342. Van Wilgen, B.W., M.P. de Wit, H.J. Anderson, D.C. Le Maitre, I.M. Kotze, S. Ndala, B. Brown, and M.B. Rapholo, 2004. Costs and benefits of biological control of invasive alien plants: case studies from South Africa. South African Journal of Science 100, 113-122. Van Wilgen B.W., D.C. Le Maitre, D.M. Richardson., C. Marais, and D. Magadlela, 2001. The economic consequences of alien plant invasions: examples of impacts and approaches to sustainable Management in South Africa. Environment, Development and Sustainability 3, 145–168. Van Wilgen, B.W., P.R. Little, A.H.M. Gorgens, T.A. Willems, and C. Marais, 1997.
The
sustainable development of water resources: History, financial costs, and benefits of alien plant control programmes. South African Journal of Science 93, 404-411. Versfeld, D.B., D.C. Le Maitre, and R.A. Chapman, 1998.
Invading alien plants and water
resources in South Africa: A preliminary assessment. Water Research Commission, Technical Report No. TT 99/98, Pretoria. Wannenburgh, A., 2009. Personal communication. Assistant director: Decision support, Working for Water (National office): Department of Water Affairs and Forestry. Woodward, I., 1987. Climate and plant distribution. Cambridge: Cambridge University Press. Working for Water, 2004.
Working for Water 2003/4 Annual Report.
Working for Water
programme, Department of Water Affairs and Forestry, Cape Town. Xu, Z-B., R. Zhang, and W-F. Jing, 2009. When does online BP training converge?
IEEE
Transactions on Neural Networks 20 (10), 1529-1539. Yuan, J., and Z. Niu, 2008.
Evaluation of atmospheric correction using FLAASH.
IEEE
Conference proceedings of the International Workshop on Earth Observation and Remote Sensing Applications. Beijing, China 30 June-2 July 2008. 119
120
Appendix A: Systematic prioritisation matrix
Figure A1: Systematic species‘ prioritisation matrix.
121
Figure A2: Systematic species‘ prioritisation matrix key.
122
Appendix B: Statistical measures utilised in the study Statistical measure Kappa
Equation
Description
k
N
k
Cii
Nri Nci
i 1
i 1 k
N2
Measure of agreement between predicted and observed values, while correcting for chance agreement between these values.
Nri Nci i 1
Overall accuracy
Overall accuracy
User accuracy
User accuracy
Producer accuracy
Producer accuracy
1 N
k
Cii i 1
Cii Nri Cii Nci
Total number of correctly classified pixels divided by the total number of reference pixels. Number of correctly classified pixels of class i, divided by the total number of pixels classified in class i. Number of correctly classified pixels of class i, divided by the te number of training set pixels used for class i. (After Lillesand et al., 2008; Liu and Mason, 2009)
Appendix C: Protocol and Algorithm DVD
Appendix D: Detail A0 Maps of the Selected IAP Species Distribution Hardcopies and DVD
123
