Urban Land Use Classification Using Street View Images Based on Deep Transfer Network

Abstract

Urban land use maps are significant references for urban planning and environmental research. Contrary to land cover mapping, it is generally impossible using overhead imagery since it requires close-up views. Street view images capture the surrounding scenes along streets and represent urban land information objectively. In this work, we proposed an automatic classification model using street view images for urban land use classification. We utilize high-level semantic image features extracted adopting a deep transfer network, which includes three fully-connected layers. Geographic information was applied to mask out land use parcel and to associate the corresponding street view images. The approach allows us to achieve 61.8% accuracy on a challenging six class land use mapping problem. The assessment results show that the proposed approach has potential on land use classification. Since the street view images are publicly accessible and supply a variety of APIs for downloading, the presented approach would provide an effective way for urban-related research in future.