CWmamba: Leveraging CNN-Mamba Fusion for Enhanced Change Detection in Remote Sensing Images
Yingchao Liu, Guangliang Cheng, Qingling Sun, Chunpeng Tian, Lukun Wang
Abstract
Remote sensing image change detection is crucial for urban construction and environmental monitoring. Recent advancements have seen convolutional neural networks (CNNs) and transformer structures increasingly applied in this domain. However, CNNs struggle with long-distance feature capture, while transformers suffer from high computational demands. Moreover, the inherently high resolution of remote sensing images and their susceptibility to natural conditions complicate feature extraction and processing, thereby hindering accurate change detection. The introduction of the mamba structure has significantly mitigated the issue of long-distance feature extraction. This letter introduces a model that integrates CNN and Mamba, named CWmamba, which employs a novel architecture combining mamba blocks and a CNN-based feature extraction block (BCGF) to process dual-temporal images. In the encoding phase, CWmamba utilizes the mamba blocks for global feature integration and the BCGF module for local feature enhancement. The decoding phase involves the fusion of multilevel features to augment the model’s expressive capability. The results of CWmamba on three datasets, SYSU-CD, LEVIR-CD+, and S2Looking, demonstrate its effectiveness, with F1 scores of 84.33%, 87.21%, and 67.93%, respectively.