Local climate zone mapping as remote sensing scene classification using deep learning: A case study of metropolitan China

TLDR: It is argued that LCZ mapping should be considered as a scene classification task to fully exploit the environmental context and more advanced domain adaptation methods should be applied in this application.

Abstract: China, with the world’s largest population, has gone through rapid development in the last forty years and now has over 800 million urban citizens. Although urbanization leads to great social and economic progress, they may be confronted with other issues, including extra heat and air pollution. Local climate zone (LCZ), a new concept developed for urban heat island research, provides a standard classification system for the urban environment. LCZs are defined by the context of the urban environment; the minimum diameter of an LCZ is expected to be 400–1,000 m so that it can have a valid effect on the urban climate. However, most existing methods (e.g., the WUDAPT method) regard this task as pixel-based classification, neglecting the spatial information. In this study, we argue that LCZ mapping should be considered as a scene classification task to fully exploit the environmental context. Fifteen cities covering 138 million population in three economic regions of China are selected as the study area. Sentinel-2 multispectral data with a 10 m spatial resolution are used to classify LCZs. A deep convolutional neural network composed of residual learning and the Squeeze-and-Excitation block, namely the LCZNet, is proposed. We obtained an overall accuracy of 88.61% by using a large image (48 × 48 corresponding to 480 × 480 m 2 ) as the representation of an LCZ, 7.5% higher than that using a small image representation (10 × 10) and nearly 20% higher than that obtained by the standard WUDAPT method. Image sizes from 32 × 32 to 64 × 64 were found suitable for LCZ mapping, while a deeper network achieved better classification with larger inputs. Compared with natural classes, urban classes benefited more from a large input size, as it can exploit the environment context of urban areas. The combined use of the training data from all three regions led to the best classification, but the transfer of LCZ models cannot achieve satisfactory results due to the domain shift. More advanced domain adaptation methods should be applied in this application.