A Complex-Valued and Loss Optimized BiSeNet for PolSAR Image Semantic Segmentation

The semantic segmentation of polarimetric synthetic aperture radar (PolSAR) image is an end-to-end method to classify the target region of PolSAR pixel by pixel. Currently, most classical neural networks used for PolSAR image classification directly use real-valued networks, which makes the loss of the phase information and the relationship between amplitude and phase information of PolSAR information, leading unsatisfactory classification results. Meanwhile, there is often an issue of uneven sample distribution in PolSAR images, which can affect the training process and reduce the effectiveness of classification. As a representative of lightweight semantic segmentation, BiSeNet can achieve good classification results in optical image segmentation. In this article, a complex-valued BiSeNet with lovász-softmax loss optimization (CV-L-BiSeNet) is proposed. Compared with original BiSeNet, CV-L-BiSeNet not only makes full use of the phase information in PolSAR data, but also reduces the impact of uneven sample classification by lovász-softmax loss. Experiments on San Francisco and Flevoland show that CV-L-BiSeNet obtains higher MPA (mean pixel accuracy), OA (overall accuracy), MIOU (mean intersection over union) and Kappa than other related semantic segmentation networks.