Accurate local climate zone (LCZ) classification is essential for urban climate studies, environmental monitoring, and sustainable city planning. Recent advances in deep learning (DL) have significantly improved LCZ mapping, but challenges remain in capturing spatial position features and distinguishing spectrally similar land cover types. This letter proposes multiscale coordinate attention-based multistream LCZ network (MSCA-MSLCZNet), a multistream DL framework integrating multiscale feature processing, attention mechanisms, and rule-based refinement to enhance LCZ classification performance. The model is evaluated on the So2Sat LCZ42 dataset, outperforming baseline methods in overall accuracy (OA), OA of built-up classes (OAbu), OA of natural classes (OAn), and Kappa. Further validation on Milan LCZ mapping confirms its generalization capability, demonstrating strong classification performance in built-up areas and improved urban structure delineation. Comparative experiments highlight the model’s ability to better differentiate urban structures and built-up zones from natural landscapes. MSCA-MSLCZNet proves effective for large-scale LCZ mapping, offering improved classification accuracy and adaptability to diverse geographic regions.
Enhancing Local Climate Zone Classification With MSCA-MSLCZNet: A Multistream Deep Learning Approach
Liu X.;Russo L.;Li W.;Gamba P.
2025-01-01
Abstract
Accurate local climate zone (LCZ) classification is essential for urban climate studies, environmental monitoring, and sustainable city planning. Recent advances in deep learning (DL) have significantly improved LCZ mapping, but challenges remain in capturing spatial position features and distinguishing spectrally similar land cover types. This letter proposes multiscale coordinate attention-based multistream LCZ network (MSCA-MSLCZNet), a multistream DL framework integrating multiscale feature processing, attention mechanisms, and rule-based refinement to enhance LCZ classification performance. The model is evaluated on the So2Sat LCZ42 dataset, outperforming baseline methods in overall accuracy (OA), OA of built-up classes (OAbu), OA of natural classes (OAn), and Kappa. Further validation on Milan LCZ mapping confirms its generalization capability, demonstrating strong classification performance in built-up areas and improved urban structure delineation. Comparative experiments highlight the model’s ability to better differentiate urban structures and built-up zones from natural landscapes. MSCA-MSLCZNet proves effective for large-scale LCZ mapping, offering improved classification accuracy and adaptability to diverse geographic regions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
