Building change detection (BCD) in remote sensing images plays a pivotal role in urban planning, disaster assessment, and land monitoring. However, traditional methods heavily rely on manual feature design, which hinders their practical applicability. Although deep learning based approaches have achieved remarkable progress in BCD, due to large variations in image size or object scales, convolutional neural networks (CNN) struggle to capture long-range dependencies and transformers are constrained by high computational complexity. To address these challenges, this paper proposes a multiscale CNN-Mamba hybrid network (MCMHNet). The network employs a CNN encoder to extract local change features from images, followed by a Mamba decoder embedded with an SENet attention mechanism to aggregate global information, thereby achieving efficient fusion of local-global representations. Furthermore, a multiscale change information capture module (MCICM) is designed, which significantly enhances the network's ability to extract multiscale building change features. Experimental results on WHU-CD and LEVIR-CD+ datasets, demonstrate that MCMHNet achieves 96.32% and 90.12% for Precision, 92.93% and 88.65% for Recall, 94.59% and 89.38% for F1- Score, 89.74% and 80.80% for IoU, respectively, outperforming state-of-the-art methods. Moreover, the training and testing time of MCMHNet is in the middle of the range and not much different from other models.
Wang et al. (Fri,) studied this question.