Land Cover Classification with GF-3 Polarimetric Synthetic Aperture Radar Data by Random Forest Classifier and Fast Super-Pixel Segmentation.

Abstract

Chinese Gaofen-3 (GF-3), a vital satellite for high-resolution earth observation, was the first C-band polarimetric synthetic aperture radar (SAR) launched in China with a resolution of up to one meter. Polarimetric SAR can obtain the complete physical scattering mechanisms of targets, thereby having the potential to differentiate objects. In this paper, several classification methods are briefly summarized and the types of features that should be chosen during classification are discussed. A pre-classification step is introduced to reduce the workload of precise labeling. The Random Forest classifier, which performs well for many other classification tasks, is used for the initial land cover classification. Then, based on a polarimetric constant false-alarm rate (CFAR) edge detector, a fast super-pixel generation method for polarimetric SAR image is proposed, which does not require the adjustment of parameters in advance. Following that, majority vote is conducted on the initial classification result based on the super-pixels, so that the classification result can be optimized to better meet the mapping requirements. The experimental results based on GF-3 polarimetric SAR data verify the effectiveness of proposed procedure and demonstrate that GF-3 data has excellent performance in land cover classification.