Detection of building structures from single polarized TerraSAR-X data

Kurzfassung

This study aims at an area-wide detection of the building structure of settlements from individual, single-polarized TerraSAR-X (TSX) intensity datasets recorded in stripmap mode. Due to SAR side-looking acquisition, the building-related information is located in areas which do spatially not exactly correspond with the true location of the buildings. To perform a supervised classification approach we at first create a mask of areas which are affected by scattering from the buildings based on reference datasets of the building footprints with their respective height by considering the viewing geometry of the TSX data. The generated mask is used in the following to randomly extract training samples in order to determine the relationships between the SAR data and the class membership. For the classification of the areas carrying the building-related information we utilize a random forest algorithm. As input features for classification we compare the suitability of the Grey Level Co-occurrence Matrix based textures measures according to Haralick, Mathematical Morphology and Spatial Autocorrelation texture measures. These features are calculated from TSX data using a pixel-based multiple-scale moving window approach. For each texture feature set and each moving window width the relationship to the class membership is modeled on the basis of the extracted training samples. The different models are used in the following to perform different classification runs of the entire TSX dataset. With the described approach we achieve overall classification accuracies of up to 78 %. The influence of the simultaneous usage of input texture features calculated with different window widths on the classification accuracy is of the same magnitude as the influence of the usage of the different texture feature sets.