Remote Sensing Image Semantic Segmentation Utilizing Geographical Element Association Features

Le résumé fourni par la source

Semantic segmentation of high-resolution remote sensing images remains challenging due to complex spatial structures, fine-grained category variations, and hierarchical semantic dependencies. Existing deep learning models often treat semantic categories as independent, flat labels, which leads to inconsistent predictions across hierarchical levels (e.g., a pixel predicted as Tree but not as Vegetation) and weak generalization in heterogeneous landscapes. To address these issues, we propose HAG Net (Hierarchical Attention Gate Multi-Residual UNet), a novel framework that explicitly models hierarchical semantics and enforces cross-level consistency. First, a multi-residual encoder–decoder backbone with hierarchical attention gates (HAGs) enhances multi-scale representation while filtering irrelevant background noise. Second, a Mixture-of-Head (MoH) attention module enables bidirectional semantic interaction between coarse- and fine-level features, mitigating error propagation caused by unidirectional designs. Third, a Hierarchical Interaction (HI) Loss introduces a dynamic category interaction matrix to adaptively constrain predictions, ensuring consistency across levels. Extensive experiments on two large-scale datasets, GID (5- and 15-class) and Ascend Cup 2020 (8- and 17-class), demonstrate that HAG Net consistently outperforms state-of-the-art methods, including DeepLabv3+, CGGLNet, and MAE-BG. Specifically, HAG Net achieves up to 4.5 % improvement in FWIoU and significantly enhances per-class accuracy for small and structurally complex objects. These results confirm the effectiveness of incorporating hierarchical semantics into segmentation networks and highlight the potential of HAG Net for large-scale land cover mapping, ecological monitoring, and urban planning applications.The source code is available at: https://github.com/yangruiqi-kiki/HAG.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.