Deep learning-based local climate zone classification using Sentinel-1 SAR and Sentinel-2 multispectral imagery

As a newly developed classification system, the LCZ scheme provides research framework for Urban Heat Island (UHI) studies and standardizes worldwide urban temperature observations. With growing popularity of deep learning, learning-based approaches have shown great potential in mapping. Three major cities China are selected as study areas. In this study, we design convolutional neural network architecture, named Residual combined Squeeze-and-Excitation Non-local Network (RSNNet), that consists (SE) block non-local to classify using freely available Sentinel-1 SAR Sentinel-2 multispectral imagery. Overall Accuracy (OA) 0.9202, 0.9524 0.9004 three obtained by applying RSNNet training data individual city, OA 0.9328 is with from all cities. outperforms other popular Convolutional Neural Networks (CNNs) terms mapping accuracy. We further series experiments investigate effect different characteristics on performance The results suggest combination can improve accuracy classification. proposed achieves an 0.9425 when integrating decomposed components images, 2.44% higher than images alone.