MULTI-SCALE AND COORDINATE ATTENTION-BASED CNN FOR LOCAL CLIMATE ZONE CLASSIFICATION

Local climate zone (LCZ) classification plays a crucial role in urban climate studies, providing detailed descriptions of urban forms and functions. Traditional machine learning methods, such as the random forest classifier used in the World Urban Database and Portal Tool (WUDAPT), struggle to achieve high accuracy due to their limitations in capturing spatial and multi-scale features. In this study, we propose MSCA-Net, a CNN framework that integrates multi-scale feature extraction, residual convolutional block attention module, and coordinate attention mechanisms to enhance LCZ classification. The model effectively captures both complex positional features and spatial contextual information, enabling accurate classification of heterogeneous urban and natural landscapes. Experimental results on the So2Sat LCZ42 dataset demonstrate the superior performance of MSCA-Net compared to benchmark models, achieving improved accuracy and more detailed LCZ mapping. These findings highlight the potential of MSCA-Net for enhancing LCZ classification and contributing to global urban climate studies.