Identification of built-up areas from TerraSAR-X data, an object-oriented approach.

Due to the continuous urbanization in developing and emerging countries more and more space is converted to urban area. Even in developed countries such as Germany, for instance, the yearly conversion of natural or agricultural space into residental, industrial or transport areas frequently exceeds 100 ha. Thus, urban environments represent one of the most dynamic regions on earth. In order to obtain more frequently updated and spatially detailed data on urban phenomena remote sensing satellites can serve as a valuable instrument. With the high spatial resolution of up to one meter, the ability to record data independently of weather or daytime and a revisit time of up to 3 days the new german radar satellite TerraSAR-X provides promising potential for collecting data on the urban morphology and dynamics.
The goal of this study was the development of a temporally and spatially robust approach towards an automated identification of built-up areas based on high resolution TerraSAR-X imagery. In view of the limited significance of single-polarized radar intensity data, the analysis of textural, contextual and hierarchical characteristics was a key issue. The image classification was based on an object-oriented analysis approach that provides lots of functionality to utilize textural, hierarchical or semantic information. Two urban environments was analysed in this study: the Rhine-Neckar region and the city of Munich in Germany.
In order to assure a robust object-oriented classification of the image an optimal segmentation was realized by an adaptive optimization procedure that improves image segmentation within the software package Definiens Developer by means of an iterative classification procedure. Since the optimized segmentation algorithm provides the best results when processing filtered SAR data, we initially filtered the TerraSAR-X intensity imagery by means of a novel speckle suppression algorithm. This filter significantly removes speckle noise while true structures, texture and contour information are preserved. Subsequently, the procedure for optimized image segmentation was performed in order to facilitate both an immediate tuning of the segmentation settings and an improvement of segmentation accuracy. The optimized image segmentation resulted in the generation of a single segmentation level (Level 2) in which all spectrally distinguishable entities are represented by individual segments. Thus, this segmentation provides a single, optimized segmentation level featuring large segments in homogeneous areas whereas small-scale structures and heterogeneous regions are represented by small objects.
Since the optimization procedure is solely based on spectral characteristics, heterogeneous regions - e.g., built-up areas - are subdivided into their spectral and structural primitives. Based on these primitives it is not possible to calculate meaningful textural features. Hence, the optimized object level was supplemented by a second, superior object level (Level 3) that featured large-area segments in order to enable both the calculation of meaningful texture measures and the utilization of hierarchical features. Finally, a third level (Level 1) with very small objects was created underneath the optimized level. This level facilitated the correction of classification errors in the subsequent image classification step that arose from an inaccurate spatial representation of image structures at the optimized object level.
Based on this pre-processing the final step of the automated extraction of built-up areas could more efficiently be focused on intensity-related, textural and contextual characteristics. The most significant characteristics of built-up areas in SAR images are the enormous heterogeneity of both backscatter as well as size and shape of urban structures in combination with an accumulation of bright punctual or linear scatterers. Hence, the rule base for the automated identification of built-up areas was focused on the analysis of these two characteristics. Since the information content of the single-polarized radar intensity imagery is comparably limited the zonal heterogeneity as well as the identification of scatterers have to be based on textural, contextual and hierarchical features. Therefore the classifying starts with the identification of built-up areas at the optimized Level 2. The classification was mainly based on the analysis of intensity information of the filtered image and texture information which was provided as a separate layer by the novel speckle suppression filter described above. Subsequent to the identification of the potential built-up areas the resulting classification was locally optimized by means of the small segments provided by Level 1.
The presented approach towards the detection of urban areas by means of TerraSAR-X data achieves accuracies around 85% for all test sites – an indicator for the high potential of TerraSAR-X data in terms of an identification of built-up areas.