Feature selection for urban impervious surfaces estimation using optical and SAR images

Urban impervious surfaces, such as transport related land (e.g., roads, streets, and parking lots) and building roof tops (commercial, residential, and industrial areas), have been widely recognized as important indicator for urban environments. Numerous methods have been proposed to estimate impervious surfaces from remotely sensed images. However, most of these approaches were proposed with optical remote sensing images, and accurate estimation of impervious surfaces remains challenging due to the diversity of urban land covers. This study presents the effort to synergistically combine optical and SAR data to improve the mapping of impervious surfaces using the Random Forest (RF). The Multilayer Perceptron, Support Vector Machine, and RF are compared for impervious surfaces mapping with the single use of optical image and with the combined optical and SAR images. Experiment shows some interesting results: 1) synergistic use of SPOT-5 and TerraSAR-X images produced more accurate classification of impervious surface mapping, no matter what combinations of features are used; 2) The SAN-based features appeared to provide effective complementary information to the conventional GLCM-based features for the classification, increasing the accuracy by about 0.6% by using supervised classifiers; 3) SVM and RF tended to be superior to MLP for the fusion of SPOT-5 and TerraSAR-X images for LULC classification and ISE. RF is better for the LULC classification as it better handled the spectral confusion among sub types of impervious and non-impervious land covers, while SVM appeared more stable before and after the combination of sub land cover types, and thus is more suitable for the ISE with the classification strategies of mapping impervious surface.