SAM-Mamba: A Two-Stage Change Detection Network Combining the Adapting Segment Anything and Mamba Models

Abstract

Remote sensing change detection (RSCD) plays a crucial role in environmental monitoring, resource management, and land planning. However, remote sensing images typically involve multiple surface types, and the similarities and differences between these types make change detection (CD) in complex scenes even more challenging. Therefore, extracting and integrating effective features in complex scenes becomes a key issue in RSCD. In recent years, SAM2 and Mamba have made a significant impact in the field of computer vision, but further research is needed in the remote sensing domain to fully leverage their advantages. Thus, this article proposes a two-stage CD network named SAM-Mamba. Specifically, we use SAM2 as the encoder to extract key features from remote sensing images and perform preliminary decoding to generate mixed prompts, which are then passed to the guiding encoder (prompt encoder) to generate accurate and stable feature boundaries. These boundaries are then fused with the key features and passed to the decoder. In the decoding stage, we integrate the Mamba model, which is designed to efficiently model long sequences and capture spatiotemporal correlations. Furthermore, to enhance the ability to capture complex changes, we design the frequency-guided feature refinement module and the dynamic feature fusion module. Our SAM-Mamba outperforms other state-of-the-art methods on four datasets (e.g., Intersection over Union and F1-score are improved by 3.40%/2.96%, 1.02%/1.07%, and 2.52%/2.42% on LEVIR-CD+, WHU-CD and SYSU-CD datasets, respectively). The code of SAM-Mamba are publicly released at https://github.com/lytwxyyyds/Sam-Mamba.