REMOTE SENSING DATA AUGMENTATION THROUGH ADVERSARIAL TRAINING

In this paper, a Generative Adversarial Network(GAN) is proposed for data augmentation of remote sensing images abstracted from Jiangsu province in China, i.e., D-sGAN(Deeply-supervised GAN). At First, to modulate the layer activations, a down-sampling scheme is designed based on the segmentation map. Then, the architecture of the generator is UNet++ with the proposed down-sampling module. Next, the generator of this net is deeply supervised by the discriminator using deep Convolutional Neural Network(CNN). This paper further proved that the proposed down-sampling module and the dense connection characteristics of UNet++ are significantly beneficial to the retention of semantic information of remote sensing images. Numerical results demonstrated that the images generated by D-sGAN could be used to improve accuracy of the segmentation network, with a better Fully Convolutional Networks Score(FCN-Score) compared to the GoGAN, SimGAN and CycleGAN models.