Remote Sensing Image Semantic Segmentation Based on Cascaded Transformer

Abstract

High-resolution (HR) remote sensing image semantic segmentation plays an important role in Earth's surface. Despite the rapid development of remote sensing image semantic segmentation methods, however, the land objects types of HR remote sensing images are complex, and the features are difficult to be extracted, the existing deep learning methods are difficult to obtain enough effective features, there is still room for further improvement in feature representation ability. In respect of the issues above, We propose a novel global and local features aggregated network to simultaneously exploit boundary information and capture hierarchical semantic information for Remote sensing image semantic segmentation. In addition, a novel loss module is designed according to Generative Adversarial Network (GAN) to enhancing the feature representation capability of the model in multi-class segmentation. In the meanwhile, the performance of CTrans_Net is verified on three public datasets, and good results are obtained. The experimental code and more results are shared at https://github.com/Mr-Wangfl/CTrans_Net .