Integrating H-A-α with fully convolutional networks for fully PolSAR classification

Classification in remote sensing, similar to semantic segmentation in computer vision, is aimed to assign a label to each pixel in images to indicate which class it belongs to. Fully convolutional networks (FCN), one of semantic segmentation methods, is proposed to tackle this problem in fully PolSAR images in this paper. To exploit the polarimetric information in PolSAR images, H-A-α polarimetric decomposition is integrated with FCN. PolSAR images acquired by Gaofen-3, China's SAR satellite, in the C-band with a spatial resolution of 1 meter are utilized. Three variations of FCN, i.e., FCN-32s, FCN-16s, and FCN-8s, and SVM are trained and validated. Experimental results reveal that the both user and product accuracy of the three FCN architectures is more than 2% higher than support vector machine (SVM) for water pixels, 16% higher for vegetation, and 24% higher for the building study areas in the whole image. Besides, the three architectures of FCN are 75 times faster than SVM.