Synergistic Use of Optical and Dual-Polarized SAR Data With Multiple Kernel Learning for Urban Impervious Surface Mapping

Abstract

Accurate mapping of impervious surface distribution is important but challenging. Integrating optical and SAR data to improve urban impervious surface estimation has recently shown promising performance. Further investigation and development on this multisensory approach are conducted in this study. A novel multiple kernel learning (MKL) framework is proposed to integrate heterogeneous features from Landsat-8 and Sentinel-1A data effectively. A linearly weighted combination of basic kernels built using each group of features is learned as the optimal kernel, while the hyperparameters and the weight of each basic kernel are determined simultaneously by using the differential evolution algorithm. Then, the optimal kernel is embedded into the support vector regression algorithm, and the impervious surface abundance of the study area is estimated by applying the developed multiple kernel support vector regression (MKSVR) model. The impervious surface ground truth at a subpixel level is derived from a high-resolution image by means of object-oriented classification. The experimental results indicate that the synergistic use of optical and dual-pol SAR data by employing MKSVR achieves a noteworthy improvement for impervious surface estimation compared to that using optical image alone, the root mean square error is decreased by 4.30%, and the coefficient of determination ( R 2) is increased by 9.47%, and that the incorporation of optical and SAR does not guarantee the improved performance, simply stacking all features of multisource data into a vector is not a good choice, and the MKL is a powerful tool to apply as demonstrated by the experiments conducted in this study.