CMNet: A CNN–Mamba Network for Change Detection With Similarity Orientation and Difference Perception

Change detection (CD) is a hot research topic in the field of remote sensing (RS), using Convolutional neural networks (CNN) and Transformer for CD tasks is the mainstream option currently. However, the limited perception field of CNN and the computational complexity of Transformer make it challenging to achieve satisfactory results in processing high-resolution RS images. In addition, many State-of-The-Art (SoTA) approaches ignore the cross-level features’ inter-temporal correlation and contextual information, which leads to inefficient feature fusion and poor perception of changed regions, resulting in ambiguous or missing detection results. To alleviate these issues, we propose a novel CNN-Mamba network with similarity orientation and difference perception (CMNet). CMNet combines Mamba and CNN in CD tasks by reciprocally coaching CNN’s local features with Mamba’s global features to learn potential semantic representations. To more effectively adapt to CD, we propose Conv-Mamba Interaction Module (CMIM), Similarity Orientation Module (SOM), and Difference Perception Module (DPM). Initially, CMIM is designed to concern the interaction of CNN, Mamba backbone, which effectively captures cross-level spatial-temporal correlations and contextual information through reciprocal learning of local-global features. Subsequently, the different hierarchical semantic features extracted by the encoder are entered into SOM for fusion. By calculating the cosine similarity between the center pixel and neighboring pixels, the similarity spatial offsets for guiding feature sampling are dynamically generated. Then, the low-resolution features are spatially aligned and up-sampled according to the similarity offsets, achieving cross-scale feature fusion. Finally, DPM further captures the changed region through the combination of cross-scale difference calculation, producing discriminative change feature representations. We evaluated the efficiency of CMNet by conducting experiments on three popular datasets. The results indicate that our method achieves SoTA accuracy while maintaining lightweight parameters. The code will be available soon at https://github.com/Jscript10/CMNet.