MIFNet: Mamba-Based Information Fusion Network for Remote Sensing Change Detection

Transformers with long-distance modeling capabilities have been widely applied in remote sensing change detection. However, the Transformer architecture introduces quadratic computational complexity, particularly in large-scale datasets. Recently, Mamba blocks based on State Space Models (SSMs) have been proposed to establish long-range feature interactions with linear complexity. However, conventional Mamba-based models often neglect the importance of local information in remote sensing images, which limits their accuracy. To address these problems, we propose a Mamba-based Information Fusion Network (MIFNet) for change detection. The backbone of MIFNet employs Mamba blocks to extract global features, while convolutional modules are integrated to complement local receptive fields and capture fine-grained details. To better fuse the bi-temporal feature maps, we design a novel Bi-Temporal Information Fusion Module (BIFM), which captures relative positional relationships between features to refine spatial information. We evaluate the performance of MIFNet on the GVLM-CD dataset, a remote sensing dataset focusing on landslide detection caused by natural disasters. Experimental results demonstrate that MIFNet achieves IoU score of 0.7703 and an F1 score of 0.8705 on GVLM-CD, outperforming compared methods.