PolSAR Image Semantic Segmentation Based on Deep Transfer Learning—Realizing Smooth Classification With Small Training Sets

Suffering from speckle noise and complex scattering phenomena, classification results of SAR images are usually noisy and shattered, which makes them difficult to use in practical applications. Deep-learning-based semantic segmentation realizes segmentation and categorization at the same time, and thus can obtain smooth and fine-grained classification maps. However, this kind of methods require large data sets with pixel-wise categorical annotations, which are time consuming and tedious to retrieve. Compared with photographs and optical remote sensing images, manually annotating SAR data is even harder, which results in a delay of using relevant techniques in this field. In this letter, a new data set is proposed to support semantic segmentation for high-resolution PolSAR images. Limited by the aforementioned problems, the data set is only a small one with 50 image patches. Therefore, two transfer learning strategies are proposed, which adopt the fully convolutional network (FCN) and U-net architecture, respectively, and use distinct pretraining data sets to adapt to different situations. The experiments demonstrate the good performance of both methods and a promising applicability of using small training sets. Moreover, although trained with small patches, both networks can perfectly apply on large images. The new data set and methods are hopeful to support various PolSAR applications as baselines.