Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

1

AN UPDATE SYSTEM FOR URBAN BIOTOPE MAPS BASED ON HYPERSPECTRAL REMOTE SENSING DATA Mathias Bochow, Karl Segl and Hermann Kaufmann GeoForschungsZentrum Potsdam, Section 1.4 Remote Sensing, Potsdam, Germany; {mbochow} {segl} {charly}@gfz-potsdam.de ABSTRACT Urban biotopes are of high importance for ecological urban planning. In contrast to the common practice of producing urban biotope maps – by visual interpretation of colour-infrared aerial photographs in combination with field investigations – this investigation presents an efficient system for updating existing urban biotope maps by automatic analysis of remote sensing (RS) data. For such a system a method for the classification of biotopes is needed. However, the automatic classification of urban biotopes from RS data is not a task for a pixel-based classifier. All pixels of a biotope have to be taken into account during the classification. Therefore, fuzzy logic models for biotope types are built with regard to the composition of the biotopes of different surface materials and their arrangement in the biotopes. The models consist of lists of numerical features, that yield the best separation of two biotope types, and associated membership functions. The features are calculated on hyperspectral images and on a normalized digital surface model and are able to numerically capture the characteristic differences between the biotopes of different types. The application of the developed models to the biotopes of six selected types in the 14.5 km2 test area in the city of Dresden, Germany, yields an overall accuracy of 87%. The types of the classified biotopes, the corresponding similarity values as well as additional attributes can be assigned to the biotopes in the output vector layer and compared to the former biotope map. INTRODUCTION Cities are centres of human activity. The intensive use of land in urban areas by housing, traffic or industrial areas leads to ecological impacts on the environment and to impacts on man’s living quality as well. To reduce these impacts, municipalities give great importance to ecological urban planning. Green spaces, for example, can serve several purposes: They are habitats for fauna, act as regulators of micro- and meso-climate [i] and at the same time they can be used as bioindicators for pollution. Urban biotope maps are an important information source for ecological urban planning [ii, iii]. They document the current state and quality of urban biotopes and are considered in landscape and town planning. Furthermore, they are accounted for in environmental impact assessments and in the impact regulation (regulated in the German Environmental Impact Assessment Act (UVPG) and in §§ 18, 19 of the German Federal Nature Conservation Act (BNatSchG) as well as §1 of the German Federal Building Code (BauGB), respectively). Area-wide urban biotope maps are produced by visual interpretation of colour-infrared photographs in combination with field investigations. Because this procedure is very time consuming and costly many municipalities do not update their existing biotope maps regularly. Thus, there is a need for a time- and cost-efficient update system that takes into account the rapid changes in urban areas to ensure an adequate monitoring of urban biotopes. The essential part of such a system must be a method to determine the type of urban biotopes, i.e. to classify urban biotopes. In [iv] we developed feature-based fuzzy logic models and showed their potential to classify biotopes of six selected biotope types. The build-up process of these models is based on expert knowledge and consists of two major steps that have to be done manually. Distinctive numerical features that were calculated on remotely sensed input data have to be evaluated and selected according to their ability to distinguish the biotope types from each other. They

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

2

serve as input variables for the fuzzy logic models. The second step is the determination of the range of the membership functions that was done also manually guided by feature distribution histograms of training biotopes. In [v] the build-up process of these models has been automated. The feature selection approach and the design of the membership functions is now based on a pairwise maximum likelihood (ML) classification of the biotopes of two biotope types. The automation optimizes the separation of the biotope types and makes the development of the models reproducible and producer-independent. As there are about 100 different biotope types defined in [ii] and [vi] the build-up of new models for biotope types is significantly improved. In this paper the totally automated processing chain from the input data to the output product is described including the generation of appropriate input data for the modelling (chapter “Preprocessing and pre-classification of input data”), the automated build-up process of the models (chapter “Building up models for the classification of urban biotopes”) and the utilization of the models for updating an existing biotope map (chapter “The update system”). The following section starts with a description of the test site and the analysed biotope types. BIOTOPE TYPES AND TEST SITE A biotope is a place where a specific community of species (biocoenosis) lives. It is a restricted area on earth (e.g. a certain hedge, a certain wall, a certain pond) characterized by similar ecological conditions. Biotopes are categorized into biotope types. Urban biotope types are especially defined by their anthropogenic use and they are – to speak strictly in terms of ecology – biotope complexes which always consist of the same conglomeration of biotopes types. A list of urban biotope types (we will not use the term biotope complexes in this work) usually used for a comprehensive biotope mapping in Germany is given in [ii] and [vi]. For this study six biotope types – detached and terrace house development, block development, perimeter block development, row house development, high-rise building development and public lawns - were selected to show the capability of the developed method. With the exception of the biotope type lawn (Fig. 1 c, EC), which contains mostly public, welltended lawns without or sparse tree presence, the selected biotope types are different types of residential areas. Thus, they appear to be quite similar to each other and are a big challenge for the method. Detached and terrace houses are 1 to 3 floor buildings in a dispersed arrangement with private gardens (Fig. 1 a, BA). Perimeter block development is characterized by surrounding line of buildings, with or without gaps, 2 to 8 floors (Fig. 1 c, BB_r). In the backyard open spaces, gardens, parking sites and garages can be found. In the similar looking block development type buildings are loosely or densely spread over the whole block (Fig. 1 a/b, BB_b). Both types can contain some trading areas. The row house development (Fig. 1 d, BB_z) distinguishes oneself by longish, parallel or orthogonal arranged buildings with spacious green spaces in-between. Two to eight floors are possible again. High-rise buildings are all buildings with 9 floors and up (Fig. 1 c, BC), usually accompanied with green spaces. The test site for this analysis is a 14.5 km2 area in the centre of the city of Dresden, Germany. It will be extended in the ongoing study. At present, 85 biotopes with detached and terrace house development, 87 with block development, 81 with perimeter block development, 56 with row house development, 64 with high-rise buildings and 134 with public lawns were analyzed for the development of the biotope type models.

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

3

Figure 1: Urban biotopes in an aerial image: BA = detached and terrace houses, BB_b = block development, BB_r = perimeter block development, BB_z = row houses, BC = high-rise buildings, EC = public lawns PRE-PROCESSING AND PRE-CLASSIFICATION OF INPUT DATA For the development of biotope type models and the update system three types of input data are needed: An existing biotope map which usually comes in vector format, hyperspectral images and a digital surface model (DSM). The existing biotope map can then be updated by an automatic analysis of the newer RS input data. The Umweltamt Dresden provided a biotope map of the year 1999 for the test site. It was converted to raster format for easier analysis. The DSM was collected in November 2002 by an airborne laser altimeter with a ground resolution of 2 m. A digital terrain model (DTM) with 20 m resolution was subtracted from the laser DSM to get a normalized digital surface model (nDSM) which contains object heights. The DSM was used for the atmospheric and geometric correction of the HyMap data. The nDSM serves as an additional input for calculating features. The hyperspectral datasets used in this study were collected by the Hyperspectral Mapper (HyMap) on the 1st of August 2000 and the 07th of July 2004. Due to the different years of data acquisition only biotopes where no principle change has occurred between 2000 and 2004 were analyzed in this study to show the potential of the method. The spatial resolutions of 3.5 m and 4 m comply with the high requirements for urban analyses. The atmospheric correction of the HyMap data was carried out using an in-house developed hybrid method (ACUM algorithm). It employs the radiative transfer models of MODTRAN in order to calculate at-surface reflectance. The algorithm also includes a scan angle dependent correction due to the large field of view of 61.3° and an adjacency effect (diffuse scattering) correction. The resulting spectra were further corrected using field spectra within an empirical line procedure to eliminate spectral spikes that occurred in the atmospheric absorption bands. Geometric correction was done with a parametric geocoding approach utilizing the in-flight recorded exterior orientation parameters and the DSM. The resulting RMS error of less than on pixel allows an accurate overlay of the input data. As it will be described in the following section, the modeling of biotope types is based on surface materials. The appropriate input data (material fraction layers) are produced by a classification and unmixing of the HyMap image. The processing chain which has reached a high degree of automation consists of (1) a feature-based endmember identification approach [vii, viii], followed by (2) a ML classification with a small threshold and (3) an iterative linear spectral unmixing [ix, x]. Classification and unmixing are supplemented with the nDSM. The whole process is illustrated in Fig. 2. At present, 19 different roof materials, 4 fully sealed and 4 partial sealed pavement types, 2 bare ground types, 3 water

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

4

types, 8 vegetation types and 2 types of shadow (on vegetation, not on vegetation) are implemented in the classification and unmixing process (Tab. 1). For step (1), different spectral variations for each of these classes were collected in a spectral library from 6 HyMap scenes covering the cities of Dresden and Potsdam, Germany. Based on these variations numerical features were computed (e.g. ratios, absorption depth, coefficients of polynomial fit) that minimize the spectral variability for an optimal identification. With these numerical features endmember pixels for each class are identified in the actual HyMap image and stored in an image-based spectral library. This library is used for the maximum likelihood (ML) classification in step (2) that results in seedling pixels for each class which are assumed to be pure. Step (3), the iterative linear spectral unmixing, starts at these seedling pixels and spatially grows like a region growing algorithm. For the unmixing of a pixel during an iteration step all meaningful endmember combinations with two endmembers are tested that contain the endmember(s) of the neighboring seedling pixel(s) (or of the already unmixed pixels in a prior iteration). The algorithm finishes when no more pixels are unmixed in an iteration. The resulting image consists of n layers - one layer per surface material - that contain the fractions of endmembers per pixel as gray values. HyMap data

Iterative linear spectral unmixing

Atmospheric and geometric correction

General Spectral Library of surface cover types

Corrected HyMap data

Feature-based Em identification

ML classification of spectrally pure pixels (seedlings)

Image-based Spectral Library

Remaining mixed pixels

|

Endmembers of previously identified pixels in the neighbourhood

List of overall meaningful EM-combinations

Most likely EM-combinations

identified pixels

Result

Best fitting result

Linear spectral unmixing with EM-combinations

Figure 2: Unmixing process of the hyperspectral data (detailed description above) Table 1: List of urban surface classes Category Roof materials

Fully sealed materials Partially sealed materials Bare ground Water Vegetation Shadow

Surface Classes Tiles (new), tiles (old), concrete, aluminum, zinc, copper, PVC, polyethylene, glass, Plexiglas, bitumen bright/dark/red, tar-paper, schist, vegetation, gravel, facade, one still unknown material (other) Concrete, asphalt, tartan track, synthetics Cobblestone pavement, concrete, red /dark loose chipping trails Sand, soil River, pond, pool Deciduous trees, coniferous trees, lawn, meadow, dry grass, field tilled, field untilled, fallow Falling on vegetation and non-vegetation

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

5

Based on the unmixing result eight additional layers were generated which contain the thematic class groups vegetation, trees, soils, roofs, metal roofs, tile roofs, flat roofs and traffic areas. These layers were built by summing up the fraction values of the respective classes per pixel. Thus, features for the identification of biotope types (explained in the following sections) can be computed either on the fraction layer of a single class or of a class group. Further input data for the feature computation is obtained from a segmentation of the class and class group fraction layers. This is done by thresholding the fraction layers with 0.5 (i.e. more than 50% of the pixel have to belong to the class) and by clumping adjacent pixels of a class (or class group, respectively) to a segment. Each class group is saved in an individual segmentation layer since overlapping can occur because of several classes that belong to more than one class group (e.g. deciduous trees is in the class group trees and vegetation). The segments of the classes are stored together in one layer. To sum up, four types of input data have been prepared for the calculation of features (see also Fig. 3): (1) the biotopes’ areas, (2) material fraction layers with surface classes and class groups, (3) segmentation layers with classes and class groups, (4) the nDSM.

Figure 3: Four types of input data shown for an example biotope (a, aerial image): (b) the biotope’s area, (c) unmixing layer of class „red roofing tiles“, (d) unmixing layer of class group “vegetation”, (e) class segments of class group “roofs”, (f) nDSM. BUILDING UP MODELS FOR THE CLASSIFICATION OF URBAN BIOTOPES From the remote sensing point of view urban biotopes contain different geo-objects (Fig. 4) (buildings, roads, trees, lawns etc.) that consist of different surface materials. In high-resolution remote sensing imagery urban biotopes, typically sized between 0.1 and 10 ha, comprise pixels of different materials. Consequently, urban biotopes cannot be classified pixel-by-pixel. Prior to a classification the extent of the biotopes must be known. It can be taken from an existing biotope map or derived from a street map because the biotopes of many urban biotope types correspond to street blocks. Based on the biotope boundaries, the presented approach employs an individual model for every biotope type that takes into account all pixels of a biotope during the classification. The models are built with regard to the composition of the biotopes of different surface materials and their arrangement in the biotopes. The premise is that different biotope types have different characteristics that can be seen in suitable remote sensing data and can be quantitatively assessed. As input data for the calculation of appropriate numerical features (Fig. 5, box “Features”; described in subsection “Feature Development”) serve the nDSM, the material fraction layers and the segmentation layers. Features are calculated based on all or selected pixels of each biotope

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

6

and the resulting feature values are assigned to the biotope. In the automatic feature selection approach (Fig. 5, box “ML-based Automatic Feature Selection”, described in subsection “Feature Selection”) distinctive features for the differentiation of two biotope types are selected. The separation is based on a Maximum Likelihood (ML) classification of the biotopes optimizing the average of omission and commission error. As a result one feature or a list of features is stored for each combination of two biotope types and incorporated into both biotope types’ models (subsection “Model Build-Up”).

Figure 4: A biotope (dashed outline) of the type “block development” in an aerial image and a HyMap image

Hyperspectral Images Classif ication / Unmixing Material Fractions

nDS M

Features

ML-based A utomatic Feature Selection

Feature Lists

Models of Biotope Types

Masks of Biotope Types V ector-RasterConversion Biotope Map (V ector Data)

Figure 5: Generation of feature-based fuzzy logic models for determining the type of biotopes from remote sensing data Feature Development Different biotope types have different characteristics. These characteristics can be expressed in numbers by automatic computation of numerical features. As it can be seen in the images of Fig. 1, for example, the building sizes of the type detached and terrace houses are generally smaller than those of the other biotope types. Thus, a feature that calculates the area of the building segments could help to distinguish detached and terrace houses from the others. Row houses are defined as longish buildings that stand exactly parallel or orthogonal to each other. Thus, they can be identified with a class segment feature for elongation and another one which calculates the angle between the class segments of a class (in this case the average angle between the class segments of the class group “roofs”). Biotopes of large-area and high buildings are characterized by high values in the nDSM for the class group “roofs” and by a high standard deviation of the height values of all biotope pixels.

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

7

The features have been developed on a knowledge base like it has been pointed out in the examples above. All developed features are published in [xi]. Each feature belongs to one of the four categories morphology of objects, arrangement of objects, percentage of area and neighborhood. The calculation of the features can be conducted on three different levels (Fig. 6): The entire biotope, a single class in a biotope or a single segment of a class in a biotope. For all basic features the statistical parameters minimum, maximum, mean and standard deviation are computed resulting in four new features. Additionally, some features are applied on three different parts of a biotope: The biotopes total area, an inward buffer from the biotope’s border and the biotopes interior. Again, three new features are generated from a basic feature so that the total number of developed features is about 2900.

Figure 6: Three levels of feature calculation. For each level the calculation of a feature is shown, exemplarily. Feature Selection Considering the selection process, it is evident that not every feature will improve the identification of every biotope type. The challenge is to find the set of relevant features that characterizes a biotope type best. In other words, the task is to find the set of features that makes it possible to identify the biotopes of a type A and to exclude the biotopes of other types from being classified as type A. However, finding one-against-all features is difficult. The presented algorithm performs a pairwise ML classification of the biotopes of two types at a time. In the first iteration the ML classification is done with every single feature. The feature with the best separation according to the mean of omission and commission error is selected. In the following iteration the ML classification is done with the previous selected feature in combination with each of the other features. Again, the feature with the highest improvement of separation is added to the feature list. This procedure is repeated until the separation of the two biotope types is 100%, the separation does not increase from one iteration to the next or a maximum number of iterations is reached. As a result of the fea-

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

8

ture selection process a single feature or a list of features is stored for every combination of two biotope types that optimally separates these two types. The selected features are listed in Tab. 2. Table 2: List of features for the separation of two biotope types BA vs. BB_b BA vs. BB_r BA vs. BB_z

BA vs. BC BA vs. EC BB_b vs. BB_r

BB_b vs. BB_z BB_b vs. BC BB_b vs. EC BB_r vs. BB_z

BB_r vs. BC BB_r vs. EC BB_z vs. BC BB_z vs. EC BC vs. EC

Compactness of roof segments; vegetated area at biotope border; elongation of biotopes; min. elongation of deciduous tree segments Compactness of roof segments; mean hight of roofing tiles; % of area of red roofing bitumen; max. height of shadow on veg.; max. height of red concrete Mean angle between roof segments; mean pixel fraction of bright roofing bitumen; % of area of trees; stdev of elongation of tiles (new) segments; stdev of height of dark loose shipping segments Local stdev of height (3x3 filter) Overbuilt area; compactness of soil segments Max. height of flat roofs in the biotope’s interior; number of polyethylene segments; max. height of roofing tiles in the biotope’s interior; % of area of asphalt in the biotope’s interior; mean distance between aluminum segments Vegetated area at biotope border ; clumpiness of roof pixels; % of area of polyethylene; central or peripher position of gray tiles; mean angle between roof segments Local stdev of height (3x3 filter); % of area of aluminum Local stdev of height (3x3 filter); max. height of gray tiles in the biotope’s interior Central or peripher position of deciduous trees; mean distance between metal roof segments; compactness of red loose shipping segments; mean perimeter of field segments; max. elongation of deciduous trees segments Local stdev of height (3x3 filter); mean height at biotope border for dark roofing bitumen; largest segment of roofing tiles Local stdev of height (3x3 filter); max. height of gray tiles in the biotope’s interior Local stdev of height (3x3 filter) Mean area of roof segments; perimeter of biotopes; compactness of biotopes; % of area of vegetation; % of area of flat vegetation Local stdev of height (3x3 filter)

Biotope types: BA = detached and terrace houses, BB_b = block development, BB_r = perimeter block development, BB_z = row houses, BC = high-rise buildings, EC = public lawns

Model Build-Up In common fuzzy logic applications input variables and membership functions are onedimensional. This is extended here as the feature lists which span an n-dimensional feature space can be seen as multidimensional input variables. For each biotope type an individual model is build from the feature lists. Each list of features for the separation of two biotope types is incorporated into the models of these two biotope types and serves there as an input variable. Thus, each model consists of n-1 multidimensional input variables whereby n is the total number of biotope types to distinguish. The multidimensional membership functions applied to these input variables are Bayesian probability functions. Just as common membership functions their output values (the Bayesian probability) ranges from zero to one. These output values are combined with the fuzzy MIN operator, which corresponds to a logical AND, in only one rule. The implication method is MIN and the output membership function is f(x) = x. The defuzzification is done by the smallest of maximum method. This set-up results in taking the smallest of all input fuzzy values (the Bayesian probability values) as the crisp output value of the models (Fig. 7). This value can be interpreted as a similarity value which expresses the similarity of the classified biotope to the type of the applied model.

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

Fuzzy operator: MIN

Implication: MIN

9

Defuzzification: Smallest of maximum

1.0 0.8

0.7

0.7

0

Input variable 1

Input variable 2

Input variable 3

0.7

1.0

Crisp output value

Figure 7: Schematic diagram of a biotope type model with three 2-dimensional input variables (i.e. feature lists with two features) for visualization purpose. The real models have 5 n-dimensional input variables. The application of the ML classifiers on the input variables yields a probability value for each. The smallest of these probability values is taken as the crisp output value of the model. THE UPDATE SYSTEM The fuzzy logic models are the main part of the update system. Their function is to classify urban biotopes. Applied on a certain biotope every model calculates a similarity value. The biotope type of the model calculating the highest value wins the biotope. This classification result was validated with stratified cross validation. While the feature selection procedure has been calculated on all available biotopes, they are divided into training and test biotopes for the accuracy assessment. The stratified cross validation was done with 10 equal-sized subsets. This method makes use of the limited number of available biotopes in a very effective way. In ten iterations the biotope type models are trained with nine of the subsets. Training means that the means and covariances of the ML classifiers for the input variables of the models are updated based on the biotopes of the nine subsets. The biotopes of the tenth subset are classified each time. A re-partitioning is done 10 times repeating the classification of all biotopes. Each time a biotope is classified it is recorded in a confusion matrix (Tab. 3). The achieved overall accuracy is 87% (kappa: 0.843). Table 3: Confusion matrix for the biotope classification (control biotopes in rows, assignment by the classifier in columns). The sum of rows (#) divided by ten gives the real number of biotopes per type because of the tenfold re-partitioning. BA

BB_b

BB_r

BB_z

BC

EC

Sum: % (#)

BA

87,4

2,1

5,3

5,2

0

0

100 (850)

BB_b

0,7

82,2

9,6

7,5

0

0

100 (870)

BB_r

1,5

12,2

70,2

14,9

1,2

0

100 (810)

BB_z

2,1

1,9

19,6

76,4

0

0

100 (560)

0

0

0

0

100

0

100 (640) 100 (1340)

BC EC Sum:

0

0

0

1,5

0

98,5

90,8 (772)

96,8 (842)

99,8 (808)

121,1 (678)

101,6 (650)

98,5 (1320)

Due to the architecture of the developed software the classification result comes in the raster format. This format is good for analyses but inefficient for storing and maintaining the data in a GIS. Since the vector polygons of the biotopes of the existing biotope map are input for the application of the models they now are used to bring the classification result into the vector format by an automatic overlay (Fig . 8). The suggested new biotope type, the corresponding similarity values of the old and the new biotope type as well as additional attributes such as the weighted surface sealing and the percentages of overbuild areas, grassy vegetation, tree-type vegetation and total vegetation [xii] (each of them outputted to an individual raster layer) can be assigned to the vector biotopes automatically by choosing the most frequent value in each raster layer. This method keeps

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

10

the original borders of the biotopes in contrast to the geometrical losses of a direct raster-to-vector conversion. The biotopes where a change was detected can be selected by a simple SQL query which compares the old and the new suggested biotope type. If desired they can be selectively inspected in the field by the local planer to validate the change. This significantly reduces the time to spend on periodical updates of biotope maps. In contrast to the common practice of re-mapping the entire city with non-automated, visual interpretation based methods every 10 years the proposed method enables an update cycle of one or two years with automatically processed updates and selective checks.

Figure 8: Application of the models for the classification of urban biotopes and raster-to-vector conversion of the result by automatic overlay CONCLUSION AND OUTLOOK The accurate identification and separation of biotope types from remote sensing data is a basic requirement for the development of an automated update system for existing biotope maps. The separation potential of feature-based fuzzy logic models has been demonstrated for six selected urban biotope types. It has been shown that the build-up process of these models is fully automated. The output result which is a vector biotope map with various additional attributes can be easily maintained and analyzed in any GI system. In the ongoing study the update system will be tested on datasets of different cities and additional biotope types will be implemented. ACKNOWLEDGEMENTS This work was funded by the Helmholtz Gemeinschaft in the framework of the HGF Research Network “Integrated Earth Observing System“. Further it was made possible by several flight campaigns carried out by the Deutsches Zentrum für Luft- und Raumfahrt (DLR) Oberpfaffenhofen, Germany. The Umweltamt Dresden provided the biotope map, the DSM and several additional data. REFERENCES i

Sukopp H & R Wittig, 1998. Stadtökologie. 2nd ed., chapter 10.6 (Gustav Fischer Verlag, Stuttgart ) 474 pp.

Proceedings 5th EARSeL Workshop on Imaging Spectroscopy. Bruges, Belgium, April 23-25 2007

11

ii

Arbeitsgruppe ''Methodik der Biotopkartierung im besiedelten Bereich'', 1993. Flächendeckende Biotopkartierung im besiedelten Bereich als Grundlage einer am Naturschutz orientierten Planung. Natur und Landschaft, 68(10): 491-526

iii

Sukopp H & R Wittig, 1998. Stadtökologie. 2nd ed., chapter 10.7 and 12.4 (Gustav Fischer Verlag, Stuttgart ) 474 pp.

iv

Bochow M, T Peisker, K Segl & H Kaufmann, 2006. Modelling of urban biotope types from hyperspectral imagery using a fuzzy logic approach. In: eProceedings of the 2nd workshop of the EARSeL SIG Remote Sensing of Land Use & Land Cover, edited by M Braun

v

Bochow M, K Segl & H Kaufmann, 2007. Automating the Build-Up Process of Feature-Based Fuzzy Logic Models for the Identification of Urban Biotopes from Hyperspectral Remote Sensing Data. In: eProceedings of the URBAN / URS 2007 Joint Conference

vi

Bundesamt für Naturschutz, 2002. Systematik der Biotoptypen- und Nutzungstypenkartierung (Kartieranleitung). Standard-Biotoptypen und Nutzungstypen für die CIR-Luftbild-gestützte Biotoptypen- und Nutzungstypenkartierung für die Bundesrepublik Deutschland. Schriftenreihe für Landschaftspflege und Naturschutz, 73, Bonn-Bad Godesberg

vii Segl K, M Bochow, S Roessner, H Kaufmann & U Heiden, 2006. Feature-based identification of urban endmember spectra using hyperspectral HyMap data. In: eProceedings of the 1st workshop of the EARSeL Special Interest Group Urban Remote Sensing, edited by P Hostert, S Schiefer, A Damm viii Heiden U, 2004. Analyse hyperspektraler Flugzeugscannerdaten zur ökologischen Charakterisierung städtischer Biotope. Dissertation. TU Berlin, digital publication, http://edocs.tu-berlin.de/diss/2003/heiden_uta.pdf ix

Roessner S, K Segl, U Heiden & H Kaufmann, 2001. Automated differentiation of urban surface based on airborne hyperspectral imagery. IEEE TGARS, 39(7): 1523-1532

x

Segl K, S Roessner, U Heiden & H Kaufmann, 2003. Fusion of spectral and shape features for identification of urban surface cover types using reflective and thermal hyperspectral data. ISPRS Journal of Photogrammetry and Remote Sensing, 58(1-2): 99-112

xi

Bochow M, K Segl & H Kaufmann, 2006. Potential of hyperspectral remote sensing for the monitoring of urban biotopes. In: eProceedings of the 1st workshop of the EARSeL Special Interest Group Urban Remote Sensing, edited by P Hostert, S Schiefer, A Damm

xii Heiden U, 2003. Ecological evaluation of urban biotope types using airborne hyperspectral HyMap data. In: 2nd GGRS/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, 18-22