Semantic Segmentation Of PolSAR Image Based On Deeplabv3

Synthetic Aperture Radar（SAR）,as a high-resolution microwave imaging radar,has allweather and long-distance imaging capabilities and can obtain surface information with high precision in a large area.It also plays a very important role in the field of crop yield estimation.Compared with single-polarization SAR,polarimetric SAR（Pol SAR）can obtain information on the scattering mechanism of targets.Therefore,interpretation techniques such as target classification and segmentation based on Pol SAR images have been further developed.With the development of deep learning,image semantic segmentation technology based on deep learning has achieved remarkable results.This technique has also been successfully applied to Pol SAR image semantic segmentation.However,the original deep learning-based image semantic segmentation networks are mainly proposed for optical images,so the input data of these networks are usually real-valued,and it is difficult to obtain ideal segmentation results when directly applied to Pol SAR data.To solve this problem,this paper proposes a lightweight complex-valued DeepLabv3+（L-CV-DeepLabv3+）for complex-valued Pol SAR image semantic segmentation.Further,an improved method based on complex-valued DeepLabv3+ and Generative Adersarial Network（GAN）is proposed.The details are as follows:1)A Pol SAR image semantic segmentation model based on L-CV-DeepLabv3+ is proposed.Compared with the original DeepLabv3+,the model is lightweight in terms of structure and parameters to be suitable for Pol SAR data with smaller datasets,while reducing the amount of computation;in addition,the model is a complex-valued network model,and the complex-valued operations involved are strictly established in the mathematical sense,it can make full use of the amplitude and phase information of Pol SAR data to mine richer target features.The experimental results on the Pol SAR dataset show that the proposed network has good semantic segmentation performance.2)A network model of Pol SAR image semantic segmentation based on complex-valued DeepLabv3+ and GAN is proposed.The main framework of the network model is GAN,in which the generator is composed of complex-valued DeepLabv3+,and the discriminator is composed of real multi-resolution convolutional neural networks.In the network training process,on the one hand,to make full use of the amplitude and phase information of Pol SAR data,complex-valued DeepLabv3+ is used to perform preliminary semantic segmentation of Pol SAR images;on the other hand,the GAN framework is introduced to further improve the performance of Pol SAR image semantic segmentation.The experimental results on the Pol SAR dataset show that the methods based on complex-valued DeepLabv3+ and GAN have better semantic segmentation performance than complex-valued DeepLabv3+.