A Complex-Valued PolSAR Image Segmentation Network With Lovász-Softmax Loss Optimization

Abstract

In recent years, complex-valued convolutional neural networks (CNNs) have emerged as a promising approach for polarimetric synthetic aperture radar (PolSAR) image segmentation by utilizing both amplitude and phase information in PolSAR data. This article introduces a complex-valued network for PolSAR image segmentation termed as complex-valued Lovász-softmax loss optimization synthetic aperture radar network (CV-LoSARNet), which is in fact a complex-valued Lovász-softmax loss optimization framework. The bilateral structure of CV-LoSARNet provides efficient feature extraction, while the complex-valued network adapting to PolSAR data can improve feature extraction capabilities. The introduced loss function combines both the Lovász-softmax loss and cross-entropy loss, which can improve the optimization objective of the segmentation. Comparative experiments conducted on E-SAR data and AIRSAR data demonstrate the superiority of the proposed network over the classical full CNN and the classic bilateral networks. Compared with the classic bilateral network, the CV-LoSARNet has improved the mean intersection over union and mean pixel accuracy of E-SAR data sets by 2.37% and 2.29%, for AIRSAR data sets, the improvement is 12.95% and 6.70%. Moreover, the segmentation performance of the proposed network on different polarimetric modes is discussed.