Polsar Image Crop Classification Based on Deep Residual Learning Network

Abstract

PolSAR image classification is one of the most basic applications of polarimetric synthetic aperture radar (PolSAR) data, which is of great significance to the research and subsequent application of PolSAR data. Traditional PolSAR image classification methods, mainly based on a single type of target decomposition method, the dimension of feature used in the process of PolSAR image classification process is relatively less and cannot make full use of the abundant feature of the PolSAR image, which is the one of the most essential characteristics of PolSAR data. With the development of deep learning, an amount of excellent deep learning models is proposed, such as deep brief net, AlexNet, deep residual network (ResNet) and so on. The classification method based on deep learning is proved to be better than traditional methods in the classification of optical and SAR images. This paper mainly analyzes the application of ResNet model in PolSAR image classification, the effectiveness of the method was proved by comparing the classical PolSAR image classification method. Firstly, some target decomposition methods were selected to calculate the multi-dimensional feature image. Secondly, the sample points of different land cover types were manually selected, and the multi-dimensional features were extracted to form the experimental data samples. Then, the PolSAR classification model based on ResNet was constructed, and the model parameters were adjusted dynamically according to the experimental sample data. Finally, the trained model was applied to the classification of experimental data, and the accuracy of the model was assessed by calculating the Kappa index of the classification result. In this paper, a quantitative index is proposed to calculate the ability of each feature to distinguish different land cover types, and the weak distinguishing feature was deleted to reduce the influence of classification independent features on the model and to improved classification accuracy. As for the speckle noise, the PolSAR image was preprocessed by simple linear iterative clustering (SLIC) superpixel segmentation, the experimental image was divided into a determined number of superpixel blocks, and the PolSAR image classification based on super-pixel blocks. Experimental results show that the PolSAR image classification method based on ResN et is conducive to the comprehensive utilization of multidimensional features of PolSAR image, the classification accuracy of PolSAR image is better than that of the classic classification method. The optimization of feature sets is beneficial to reduce model training time and improve the classification accuracy of PolSAR image as well. The superpixel segmentation is beneficial to reduce speckle noise and further improves the accuracy of classification.