A New Parallel Dual-Channel Fully Convolutional Network Via Semi-Supervised FCM for PolSAR Image Classification

Convolutional neural network (CNN) has achieved remarkable success in polarimetric synthetic aperture radar (PolSAR) image classification. However, the PolSAR image classification is a pixelwise prediction assignment. The disadvantages of repeated calculation, memory occupation, and inadequate labeled pixels make that CNN is less efficient when classifying PolSAR images. Compared to CNN, fully convolutional network (FCN) has obvious merits for PolSAR image classification: take arbitrary-size images as input; output a 3-D dense class map to maintain the structure of the input; settle PolSAR image classification tasks with efficient dense learning, which avoids repeated calculation and memory occupation. In view of accuracy, efficiency, and insufficient labeled pixels, our method proposes a new parallel dual-channel dilated fully convolutional network (DCDFCN) for PolSAR image classification. First, limited by insufficient labeled samples, the semi-supervised fuzzy c-means clustering algorithm is exploited to enlarge the labeled samples set. Then, to improve the density of output feature maps, we design a new FCN variant named dilated FCN (DFCN) by introducing a dilated convolution block to FCN. Finally, we use two similar DFCN frameworks in parallel with different convolutional kernels to design a new DCDFCN model, which can extract more discriminate features of PolSAR images. Experimental results on three PolSAR images demonstrate that the performance of our approach is superior to several state-of-the-art methods for PolSAR image classification.