Monitoring of Green, Open and Sealed Urban Space URBIS – EO data based support for sustainable urban development

K. Jupová, T. Bartaloš, T. Soukup

G. Moser, S. B. Serpico, V. Krylov, M. de Martino

GISAT Prague, Czech Republic [email protected], [email protected]

University of Genoa, DITEN Dept. Genoa, Italy [email protected]

N. Manzke

N. Rochard

Osnabrück University Osnabrück, Germany

ADUGA Amiens, France

Abstract — In recent decades, urban sprawl became a serious European-wide problem, not only due to total land taken, but also due to its spatial distribution patterns and quality of the land consumed. Land use efficiency is becoming a prime political objective at both European and city level, and the EU Land Communication aims to establish “zero net land take” across the EU by 2050. Land is a finite resource and therefore the sprawl has to be regulated. This can be realized through careful management of urban land, applying the concept of urban land transformation. The other main concern of sustainable city panning is the preservation and improvement of the Green Infrastructure in the cities. Only such approaches can assure sustainable development of European urban and suburban landscape as well as of quality of environment and life in the cities in the long term perspective. The EU funded URBIS project (ICT PSP 2014–17) targets these issues and focuses on investigation of open space potential in urban areas, and the opportunities for inner development, as well as on investigation of urban green systems. URBIS delivers EO based methodologies and tools to provide accurate up-to-date intelligence that is comparable across European cities to support the definition and implementation of sustainable planning and governance strategies in cities and cityregions throughout Europe. The present paper focuses on the role, within the URBIS methodologies and products, of remote sensing image analysis. Classification, feature extraction, and multitemporal analysis approaches are combined to characterize open, green, and sealed areas within large urban zones. Keywords— ESA Programme: Copernicus, Urban, Land Cover and Land Use, Sustainable Development, Green and open space, Soil sealing, Image analysis, EO data classification.

I. INTRODUCTION URBIS services are built on data acquired in frame of the Copernicus programme. This generates large number of standard open-data, Earth Observation datasets as well as standard services (Urban Atlas, High Resolution Layers (HRLs)), which can be effectively exploited to support sustainable urban and regional planning and policies. For this purpose, image analysis is applied on satellite imageries and combined with thematic information about land use provided directly by Urban Atlas layer, to obtain the

overview about the amount and distribution of urban sealed, green and open space, including potential development sites, and its dynamics. Image analysis methods helps to gather additional information from VHR imagery, which is not available in Urban Atlas (due to MMU applied), in particular about internal structure of sites, but also about distribution of green and open spaces in the city in general. These techniques, utilizing both pixel and object based image analysis (OBIA) approach, are dealing with multiple spatial, spectral and textural image characteristics. They are applied by default on SPOT5 imagery acquired for European Urban Atlas mapping, but they can be applied to higher resolution imagery as well. As a result, the inventory of potential development and green sites in urban areas, including their characteristics detectable from EO data, like level of sealing, amount and type of vegetation cover etc., has been obtained. Land cover classification maps are produced, followed by derived thematic products – Enhanced Imperviousness Map and Green and Open Space Layer. Then, based on these intermediate layers, indicators are calculated to provide information about spatial distribution inside larger analytical units like sub-city districts, and about the overall urban redevelopment potential of the city. Preparing detailed methodology, the robustness is an important factor, to assure the applicability of workflow for potential European-wide analysis in the future. The methodology is demonstrated through implementation on three pilot sites - Greater Amiens (FR), Osnabrueck (DE) and the Moravian-Silesian Region (CZ). Baseline services for the year 2012 and backdating towards 2006 have been implemented. II. IMAGE ANALYSIS BASED ON EO DATA A. Advanced image analysis methods The central product of the image analysis chain developed in URBIS is given by the land-cover classification maps, with pixel spacing of 2.5 meters. These raster products express the class membership of each pixel, on the basis of spectral and

978-1-5090-5808-2/17/$31.00 ©2017 IEEE

textural features and local spatial information, in combination with the class assignment in the Urban Atlas. Following classes are automatically detected: Buildings, Paved surfaces, Low vegetation, Tall vegetation, and Bare soil. Then, three additional classes are imported directly from the Urban Atlas thematic layer: Water surfaces, Railroads, and Road network. As the first intermediate output, a raster map reporting the thematic land cover classification map regarding classes of interest is produced.

computed on test samples disjoint from the training set; analysis of the visual quality of the mapping products; and feedback from the end-users of the project. For example, TABLE I reports the accuracies, on the test samples, of the land cover maps computed on the three pilot sites from images collected in 2012 (baseline service). High accuracies were obtained for almost all the classes, a result that confirms the effectiveness of the joint use of the aforementioned image analysis methods. The integration of pixelwise and OBIA approaches also was effective in ensuring satisfactory visual quality, although the spatial resolution of the input SPOT5 data (5-10m depending on the channel) was poorer that the nominal pixel spacing (2.5m). TABLE I CLASSIFICATION ACCURACY OF THE URBIS BASELINE LAND COVER MAP (2012) ON THE TEST SET OF EACH PILOT SITE Class buildings paved low veg bare soil tall veg AA

Figure 1 Land Cover classification map - example from Amiens (2012)

The land-cover classification product (Figure 1) is computed based on advanced state-of-the-art feature extraction, classification, multi-temporal image analysis, and domain adaptation methods. Feature extraction techniques including texture analysis (grey-level co-occurrence and semi-variogram) and vegetation indices are used to characterize spatial information [5], [7]. Supervised classifiers based on ensemble (random forest), kernel (support vector machines) [5], [7] , and deep (convolutional neural nets) learning [6] are applied to discriminate among the aforementioned classes. Domain adaptation tools are employed for the production of classification results over large areas (e.g., in case of large urban zones composed on multiple satellite acquisitions). They favor portability of trained classifiers and allow minimizing the operational requirements in terms of training data [6]. For the refinement of classification results, also OBIA methods are applied in the frame of post-processing. For the URBIS update service multi-temporal image analysis is of major importance. The availability of two (or more) multi-temporal acquisitions of the same geographical site is exploited to identify temporal transitions among the land-cover classes. In the multi-temporal case, individual classification maps are generated for the considered dates as well as the detailed change detection map highlighting temporal transitions – all in the frame of URBIS update services. The resulting land cover maps were extensively validated through: quantitative performance figures (e.g., producer accuracy, PA, user accuracy, UA, average accuracy, AA)

Amiens PA UA 0.8557 0.9975 0.9191 0.6133 0.9967 0.9175 0.8748 0.9732 0.9476 0.9379 0.9188

Osnabrueck PA UA 0.8872 0.9853 0.9395 0.815 0.9902 0.9959 0.9981 0.9633 0.9829 0.9924 0.9596

Moravian-Silesian PA UA 0.8655 0.9999 0.9868 0.5788 0.9843 0.9359 0.9502 0.9576 0.9636 0.9801 0.9501

B. Enhanced Imperviousness Map The Enhanced Imperviousness Map is a raster map with perpixel estimates of the sealed soil presence, as an index of the degree of imperviousness quantized into five levels {0%, 25%, 50%, 75%, 100%} and reported at 5m resolution (Figure 2).

Figure 2 Example of “enhanced imperviousness map” – Amiens (2012)

In the URBIS workflow, this product further serves as an input dataset for calculation of some of URBIS criteria or indicators. The adjective “enhanced” is aimed at recalling the improved spatial resolution as compared to the Copernicus Imperviousness Layer (20m resolution). This layer is derived from the land cover map by computing the per-pixel percentage of impervious surface on the 5m pixel

grid according to the land cover estimated on the 2.5m grid. Accordingly, it can be derived from either single-date or multitemporal image analysis. In the former case, it is aimed at mapping imperviousness at a given time within the baseline service. In the latter case, the goal is to capture the temporal evolution of imperviousness in between two observation times within the update service. C. Vegetation Map Identified non-sealed surfaces forms the “vegetation map”. This product presents thematic classification map regarding vegetated areas, classified into the following three classes: Tall vegetation (in the urban context - such as park areas - and in semi-urban zones), Low vegetation (including arable land), and Bare soil. Similar to the enhanced imperviousness map, it is mapped at 5m resolution from either single-date or multitemporal image analysis within the baseline and update services, respectively. III. URBIS GREEN AND OPEN SPACE LAYER The “Green and Open Space Layer” is a thematic raster dataset in spatial resolution of 2,5m, covering all sites representing green and potential development areas (PDAs) in the urban and suburban zones. Potential development areas are gaps in urban structures, both vegetated and non-vegetated, with potential for new urban development. Exploitation of such sites can help to mitigate uncontrolled urban sprawl in suburban areas

Figure 3 Example of URBIS Green and open space layer – Amiens (2012); (GS – Green Site, PDA – Potential development area)

. The layer contains following classes: Green Site – Urban, Green Site (GS) – Non-urban, Green Site – Non-urban (outside city area), PDA – Vacant or underused land, PDA – Gap in builtup area, PDA – Greenfield, and Former extraction sites covered by higher vegetation (Figure 3). These classes as well as classification rules have been defined based on users´ requirements for identification of different types of surfaces, in combination with feasibility of source data and image analysis methods. A methodology for

generation of this layer takes advantage from integration of information contained in Copernicus Urban Atlas (preselected areas) with image analysis results (land-cover classification map) to address all types of green sites and PDAs. Membership to the “Green and Open Space Layer” as complement to buildings, roads, and water bodies inside the urban area is estimated on the basis of assignment to Urban Atlas classes in combination with spectral and textural and geometrical (area) features extracted by image analysis results where further local spatial neighborhood and possible membership to a homogeneous image region and land cover class is considered including. IV. VHR DATA POTENTIAL All the analyses were performed on SPOT5 data, acquired in the frame of Copernicus programme, which were also exploited as a basis for Urban Atlas inventory mapping. SPOT5, with 5m resolution in MS mode, is reliable for green and open space detection, however, more recent satellite sensors with higher spatial and spectral resolution give significant advantage for classification. To assess the potential of these data within the URBIS framework, further experimental analysis has been conducted with World View-2 (WV-2) imagery. This analysis was in view of the foreseen future availability of WV and Pleiades data in operational mode in the frame of EU initiatives. The algorithms tested on SPOT5 data were applied to WV-2 imagery (panchromatic 0.46m, multispectral 1.84m with 8 bands, 12bpp dynamic range), covering parts of the same cities.

Figure 4 Comparison of land cover detection results using WV-2 and SPOT5 data (upper left – WV-2 false color, upper right – pixel-wise classification of full land cover (WV-2), lower left – OBIA based aggregated classification (aggregated: buildings and paved, low vegetation and bare), lower right – pixel-wise filtered classification of full LC (SPOT5)

The comparison of the results confirmed the high accuracy that could be obtained with the data from either space mission and also suggested that, as expected, the higher spatial resolution of WV-2 data is very beneficial for the precision of the detection of green and open spaces (Figure 4). These results

point out the potential of the further update and extension of the developed processing chain to incorporate higher-resolution or possibly multimodal data. V. DERIVED INDICATORS The resulting products – Enhanced Imperviousness Map and Green and Open Space Layer – than serve as a basis for calculation of indicators, which should help to evaluate and benchmark diverse analytical units, in sense of internal characteristics and spatial distribution of sealed or green and open spaces inside the unit. First, simple indicators are calculated, expressing amount and percentage of sealed, green and open space in each analytical unit (represented by fully comparable 1km grid cells, as well as by administrative units like sub-city districts followed by LAU2 and FUAs). Second, advanced urban sprawl and fragmentation indicators are calculated using algorithms developed in the EEA framework, to describe the spatial pattern of sealed, green and open areas, including connectivity. According to experiences gained from numerous projects dealing with the land cover/land use analysis, following types of metrics proved to be useful for the purpose of description of the spatial pattern of the urbanized areas: Urban sprawl metrics, Landscape fragmentation metrics. A set of urban sprawl measures taking into account spatial configuration of the urban areas (not just total amount) described in [1,2] has been applied, including following indicators: Weighted Urban Sprawl/Proliferation (WUP), degree of urban dispersion (DIS), total sprawl (TS), degree of urban permeation of the landscape (UP), Utilization Density (UD) and sprawl per capita (SPC).

applied to multispectral data from recent space missions (available in frame of Copernicus programme), are convincing tools for the detection of sealed, green and open spaces in urban and suburban areas as well as for monitoring of their development in time. Results derived from SPOT5 data were found accurate within this application. The main drawback remains the reduced quality of pansharpened supermode imagery not corresponding to original spatial resolution around 5 m. This is expected to be further significantly improved by using VHR imagery as a source data for classification. The thematic masks of sealed areas or green and open space can further serve for calculation of indicators describing spatial pattern of these types of surfaces in urban and suburban areas. Such information can serve as valuable support for decision making, development and urban area management, especially in combination with other land-use based and socio-economic indicators, as already confirmed by pilot users. For example, it can help to detect locations endangered by lack of green and open space and try to identify potentials for green area connectivity improvements. Such support can lead to significant improvement of green infrastructure, as well as improvement of the overall quality of life in the city and its hinterland. ACKNOWLEDGMENTS Special thanks to contribution of all participating project partners: Ch. Sannier, S. Delbour (SIRS), D. Ludlow (UWE), U. Ferber, K. Eckert (STADT+), N.de Lange, S. Xu, B. Albers, J.H. Haunert, A. Wichmann, M. Kada, (UOS). REFERENCES [1]

Also the landscape fragmentation measures - effective mesh size and effective mesh density - are dealing with the spatial pattern of the urban and suburban landscape. More specifically, they serve for quantification of the connectivity of nonartificial patches of land. As a consequence of urban sprawl, in particular of the linkage of built-up areas via linear infrastructure, such as roads and railroads [3],[4], fragmentation of landscapes is rising and the remaining ecological network provides less and less connectivity, which has negative impact e.g. on biodiversity. Details on fragmentation measures calculation are provided in EEAFOEN report [8].

[2]

Previously, these measures were applied on Copernicus HRL Imperviousness or Corine LC as a source data. This time, the same methodology has been applied on URBIS basic products - Enhanced Imperviousness Map and Green and Open Space Layer – to gather even more precise description of the spatial pattern of urbanized areas and also of green and open spaces inside them.

[8]

VI. DISCUSSION The results gathered in the frame of URBIS project have shown that current advanced image analysis methods, when

[3] [4] [5] [6] [7]

[9]

Jaeger, J.A.G., et al. (2009): Urban permeation of landscapes and sprawl per capita: New measures of urban sprawl, Ecological Indicators 10 (2010),$27–441. Schwick, C. et al. (2012): Improving the measurement of urban sprawl: Weighted Urban Proliferation (WUP) and its application to Switzerland, Ecological Indicators 38 (2014), 294–308. Saunders D.A., et al. (1991): Biological Consequences of Ecosystem Fragmentation: A Review, Conservation Bilology 5(1) (1991), 18-32. Forrman, R.T.T., (1995): Land mosaics — The ecology of landscapes and regions, Cambridge University Press, Cambridge/New York, 632 pp. G. Moser, V. Krylov, M. de Martino, S. B. Serpico, “The URBIS project: Vacant urban area classification and detection of changes,” Proc. of JURSE 2015, Lausanne, Switzerland, March 30 – April 1, 2015 V. Krylov, M. de Martino, G. Moser, S. B. Serpico, “Large urban zone classification on SPOT-5 imagery with convolutional neural networks,” Proc. of IEEE IGARSS 2016, Beijing, China, July 2016, pp. 1796-1799 J. A. Richards, Remote sensing digital image analysis, 5th ed., Springer, 2013 Landscape fragmentation in Europe, Joint EEA-FOEN report, EEA Report No 2/2011, ISSN 1725-9177 EEA (European Environmental Agency): http://www.eea.europa.eu/