Abstract

Urban land-use mapping is a significant yet challenging task in the field of remote sensing. Although numerous classification methods have been developed for obtaining land-use information in urban areas, the accuracy and efficiency of these methods are insufficient to meet the requirements of real-world applications such as urban planning and land management. In recent years, deep learning techniques, especially deep convolutional neural networks (DCNN), have achieved an astonishing level of performance in image classification. However, the traditional DCNN methods do not focus on multispectral remote sensing images with more than three channels, and they are limited by their training samples. In addition, these methods uniformly decompose large images into small processing units, which chop up the land-use patterns and produce land-use maps with obvious “blocks”. In this study, a semi-transfer deep convolutional neural network (STDCNN) approach is proposed to overcome these weaknesses. The proposed STDCNN has three parts: one part involves a transferred DCNN with deep architecture; another part is designed to analyze multispectral images; and the final part fuses the first two parts into a classification layer. Moreover, a skeleton-based decomposing method using street block data is devised to maintain the integrity of the land-use patterns. In two case studies, the proposed method is used to generate urban land-use maps from a WorldView-3 image of a 143 km2 area of Hong Kong and a WorldView-2 image of a 25 km2 area of Shenzhen. The results show that the proposed STDCNN obtains an overall accuracy (OA) of 91.25% and a Kappa coefficient (Kappa) of 0.903 for Hong Kong land-use classification, and an OA of 80% and a Kappa of 0.780 for Shenzhen land-use classification. In addition, due to the proposed skeleton-based decomposition method, the proposed method can produce better land-use maps for real-world urban applications.