CAMGNet: class activation map guided network for urban village extraction from remote sensing image

Urban villages exhibit complex spatial structures, ambiguous boundaries, and high heterogeneity, posing significant challenges to intelligent extraction from remote sensing images. To address the limitations of existing methods in global contextual modeling, boundary delineation, and cross-domain generalization, this study proposes a novel class activation map-guided network named CAMGNet, which integrates scene-level guidance and multi-scale boundary enhancement. CAMGNet adopts a three-branch architecture that combines scene classification encoding, semantic encoding, and semantic decoding. Grad-CAM is employed to generate class activation maps from the scene classification branch, guiding the model to focus on urban village regions of interest. In addition, a multi-scale boundary enhancement (MBE) module is designed to enhance the model's ability to extract detailed boundary information. Extensive experiments conducted on two urban village datasets demonstrate that CAMGNet consistently outperforms existing methods in terms of <i>precision</i>, <i>recall</i>, <i>F</i>1−<i>score</i>, and <i>IoU</i>. Overall, this study offers an effective solution for the automated recognition of complex urban spatial structures in high-resolution remote sensing images.