Domain Adaptation and Super-Resolution Based Bi-Directional Semantic Segmentation Method for Remote Sensing Images

Abstract

Image semantic segmentation methods based on convolutional neural network rely on supervised learning with ground truth, thus cannot be well extended to datasets that all of the data are unlabeled. Domain adaptation can solve the problem of inconsistent feature distribution between target and source domains. However, when the spatial resolution of remote sensing images in the source and target domains are not the same, those domain adaptation methods are not effective. In this paper, we propose a bi-directional semantic segmentation method based on super-resolution and domain adaption (BSSM-SRDA). With the help of generative adversarial learning, the method accomplishes semantic segmentation task from a low-resolution labelled data source domain to a high-resolution unlabelled data target domain by reducing differences in resolution and feature distribution. In addition, we propose a self-supervised learning algorithm that helps the domain discriminator to focus on those target data that has not been aligned with the source domain. The experiments demonstrate the superiority of the proposed method over other state-of-the-art methods on two remote sensing image datasets.