Optimizing the classification ability of CNN for SAR fully polarized radar data based on DCGAN

Abstract

In recent years, the deep learning network is widely used. In the field of remote sensing images, due to the high cost of image acquisition, there are still too few training samples, which greatly limits the application of deep learning in SAR data classification. This paper proposes a method that is generating simulated SAR image by generative adversarial network, and uses the image as the training data of convolutional neural network. Aiming at the impact of the simulated images generated by DCGAN’s generator on the classification of convolutional neural networks. The results show that DCGAN can fully extract the main features of the image, and the convolution model based on DCGAN can make CNN have better classification ability and get rid of the dependence on the sample size. CNN can also make full use of simulation data. Whether it is test data set or random dataset, its F1 score can obviously surpass the classification ability without DCGAN’s simulated data. In experiments with different sample numbers, the highest F1 score is 93.6479 in the dataset with DCGAN’s simulated data. In another experiment, its F1 Score reached 87.32, higher than the dataset without DCGAN’s simulated data.