中国科学院自动化研究所机构知识库

In recent years, change detection (CD) of optical remote sensing images has made remarkable progress through using deep learning. However, current CD deep learning methods are usually improved from the semantic segmentation models, and focus on enhancing the separability of changed and unchanged features. They ignore the essential characteristics of CD, i.e., different land cover changes exhibit different change magnitudes, resulting in limited accuracy and serious false alarms. To address this limitation, in this article, a category context learning-based difference refinement network (CLDRNet) based on our previous work is proposed. Considering the semantic content differences of heterogeneous land covers, a category context learning module is designed, which introduces a clustering learning procedure to generate an overall representation for each category, guiding the category context modeling. The clustering learning process is differentiable and can be integrated into the end-to-end trainable CD network, so it considers the semantic content differences from the CD perspective, thereby improving the CD performance. In addition, to address the magnitude differences of different land cover changes, a two-stage CD strategy is introduced. The two stages correspond to difference map learning and difference map refinement, aiming at ensuring high detection rates and revising false alarms, respectively. Finally, experimental results on three CD datasets verify the effectiveness of our CLDRNet in both visual and quantitative analysis.