An integrated network of semantic segmentation and change detection for remote sensing images

Currently, most of the research on change detection focuses on the traditional binary change detection (BCD), a method that only needs to focus on whether a region has changed or not. With the wide range of change detection applications, there is a trend to study the categories of changes. However, current approaches to Semantic Change Detection (SCD) typically handle semantic segmentation and change detection as distinct, isolated steps. This separation often leads to the propagation of errors from one stage to another, which in turn undermines the overall performance of the system. To address this limitation, We develop an integrated network where semantic segmentation and change detection networks (ISCN) interact through cross-temporal feature fusion, enabling simultaneously locates changed regions while maintaining semantic coherence across temporal phases. And then, we propose an adaptive class balance learning method to improve the accuracy for minority class recognition caused by the uneven distribution of changing types. Additionally, a domain adaptation module is proposed to decrease the false changes caused by the seasonal differences between two images. We test the performance of our proposed approach on a publicly available dataset, which can achieve a satisfied performance.