Synergistic Use of Optical and SAR Data with Multiple Kernel Learning for Impervious Surface Mapping

摘要

Accurate estimation of impervious surfaces is important but challenging due to the spectral confusion between different land covers. Recently, the synergistic use of optical and Synthetic Aperture Radar (SAR) data has shown advantage in improving impervious surfaces estimation. In this paper, multiple kernel learning (MKL) was employed to combine the heterogeneous features of Landsat-8 and Sentinel-1A data. Impervious surface was estimated at a sub-pixel level based on support vector regression (SVR) model. The impervious surface percentage (ISP) of training data was derived from high resolution image using the object-oriented classification approach. The results indicate that the combined use of optical and SAR by using MKL can significantly improve the estimation of impervious surface, since it not only reduces the underestimation and overestimation of ISP in urban areas, but also well separates bare soils from impervious surface. Compared with using optical image alone, the root mean square error (RMSE) is decreased by 5.5 % and the coefficient of determination (R^{2) is increased by 8.8 %.}