Enhanced Neighborhood Information Extraction for Local Climate Zone Classification by Coupling a Global Attention Mechanism
IEEE Access · 2026 · 14 · 164773-164783
DOI: 10.1109/access.2026.3745988 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yihan Li und 7 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Jiangsu Province Graduate Research and Practice Innovation Program, National College Students Innovation and Entrepreneurship Training Program.
Abstract (Originalsprache)
Accurate Local Climate Zone (LCZ) classification is essential for urban climate studies, but distinguishing built-up LCZ classes remains challenging because many categories have similar spectral responses but different neighborhood-scale morphology. To improve spatial–spectral contextual representation, this study proposes a Multi-scale Multi-level Attention Network, termed MMA-Net, by coupling a Global Attention Mechanism (GAM) with a modified ResNet50 backbone. MMA-Net integrates a multi-scale feature extraction module, an SE-enhanced backbone, and GAM-based multi-level attention branches to strengthen neighborhood information extraction from Sentinel-2 image patches. Experiments were conducted on the So2Sat LCZ42 Version 3 block-split dataset. MMA-Net achieved an OA (Overall Accuracy) of 0.90 and Kappa of 0.89, improving OA and built-up OA by 10 and 11 percentage points over the baseline, respectively. Ablation experiments confirmed the complementary roles of the multi-scale module and GAM-based multi-level attention branches. The attention comparison further showed that the GAM-based configuration achieved higher classification accuracy than the variant incorporating the Convolutional Block Attention Module (CBAM), although this improvement was accompanied by increased model complexity. Application to Suzhou, China, provided a qualitative demonstration that MMA-Net can produce spatially reasonable LCZ patterns in a heterogeneous high-density urban environment. These results indicate that GAM-based neighborhood information extraction is an effective strategy for Sentinel-2-based LCZ classification and can support more reliable urban climate analysis and climate-adaptive urban planning.
Förderung
- National Natural Science Foundation of China
- Jiangsu Province Graduate Research and Practice Innovation Program
- National College Students Innovation and Entrepreneurship Training Program
Zitieren
APA
Li, Y., Lu, X., Sun, T., Liu, Y., Zhang, W., Zhang, M., Zhang, H., & Cui, X. (2026). Enhanced Neighborhood Information Extraction for Local Climate Zone Classification by Coupling a Global Attention Mechanism. IEEE Access, 14, 164773-164783. https://doi.org/10.1109/access.2026.3745988
Vancouver
Li Y, Lu X, Sun T, Liu Y, Zhang W, Zhang M, et al. Enhanced Neighborhood Information Extraction for Local Climate Zone Classification by Coupling a Global Attention Mechanism. IEEE Access. 2026;14:164773-164783. doi:10.1109/access.2026.3745988
BibTeX
@article{Li2026,
author = {Li, Yihan and Lu, Xuehe and Sun, Tong and Liu, Yingying and Zhang, Wenwen and Zhang, Mengru and Zhang, Haidong and Cui, Xuan},
title = {Enhanced Neighborhood Information Extraction for Local Climate Zone Classification by Coupling a Global Attention Mechanism},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {164773--164783},
doi = {10.1109/access.2026.3745988}
}
EUVIMED