Extended U-Net for Satellite Image Semantic Segmentation

Abstract

In this paper, we propose an extended model of encoder-decoder structure using U-Net to improve accuracy in satellite image semantic segmentation. The U-Net is a deep learning-based semantic segmentation scheme, and research is being conducted in various applications such as road sign detection, medical image analysis, and tumor detection. The conventional U-Net suffers from losses during feature compression-expansion due to shallow structures of encoder-decoder. This creates a problem of reduced segmentation accuracy. In order to address this issue, the proposed scheme exploits an extended structure using concatenate upsampling and residual learning, which remedies the loss of information and improves the accuracy of semantic segmentation. In the experiments, segmentation was performed on various satellite images, and it was shown that the proposed U-Net was superior to the conventional counterpart.