Mapping Urban Form and Land Use with Deep Learning Techniques: A Case Study of Dongguan City, China

Precise urban form and land use data are crucial for various modeling simulations. Approaches using spectral features of remote sensing (RS) images have been widely adopted to capture land surface patterns. However, challenges arise concerning the accurate classification of built areas due to high levels of heterogeneity. An effective way to address this issue is to incorporate spatial contextual information. Deep learning techniques have recently been applied to RS image classification, and impressive results have been reported. This research undertook a case study of Dongguan City, which is a core Chinese city characterised by high density and heterogeneity. A culturally neutral classification scheme called local climate zone was adopted for land surface classification. The default random forest classification was used as a benchmark, and the moving window approach and several pretrained convolutional neural network (CNN) models were also applied for comparison. The moving window approach achieved the highest mapping accuracy with a window size of 5 x 5. Although the CNN models failed to achieve state-of-the-art results, they still exhibited an excellent performance considering the small number of input features and small input patch size. They can also reduce the salt-and-pepper phenomenon effectively.