Semantic Segmentation of Polarimetric SAR Imagery Using a Few Well-selected Training Samples

During the last decade, multi-frequency and polarimetric SAR (PolSAR) imaging has been investigated with respect to classification of terrain types, many supervised and unsupervised segmentation and classification methods for PolSAR data have been proposed. However, it is still very difficult to get a reliable and consistent scene semantic segmentation for PolSAR imagery. Recently, with the introduction of Conditional Random Field (CRF), the use of discriminative models for semantic segmentation and classification tasks has become popular [1]. CRF directly models the conditional probability which is able to incorporate a rich set of arbitrary nonindependent overlapping features of the observations. It has been shown to outperform the MRF based generative models in semantic segmentation of natural images [2]. In general, these systems require a large number of full labeled images to produce an effective segmenter. Unfortunately, it requires the labor-intensive and timeconsuming works to label every pixel in training samples. Furthermore, due to the intrinsic speckle phenomenon, sometimes only experts of SAR image interpretation are qualified for making ground truth labels. This often limits the amount of training data available, which can lead to an inferior segmentation system. In this work we will investigate the semantic labeling of PolSAR imagery based on CRF with a few well-selected training samples, and an active learning process will be employed to address the above-mentioned problem by identifying which of the unlabeled images should be labeled [3].