Compare different levels of fusion between optical and SAR data for impervious surfaces estimation

Abstract

Impervious surfaces estimation has become very significant both in urban and environment studies. Recently, the synergistic use of optical and Synthetic Aperture Radar (SAR) data has been identified as a promising approach for accurate estimation of impervious surfaces. This paper presents a comparison study over three different levels (pixel, feature, and decision level) of fusion between optical and SAR data. Detailed fusion strategies are designed at three levels, and the support vector machine is employed as the fusion method. Results indicate that pixel level fusion (Overall accuracy: 78.27%; Kappa: 0.5654) seems not appropriate for optical-SAR fusion, as it reduces the accuracy compared to the single use of optical data (Overall accuracy: 78.47%; Kappa: 0.5694). In addition, decision level fusion obtained the best accuracy (Overall accuracy: 89.87%; Kappa: 0.7974), since it not only improves the impervious surfaces in shaded areas and but also well separates bare soils.