Classification of Urban Functional Zones by Integrating Spatial Features of VHR Satellite Images and Semantic Features of POI Data

Abstract

An accurate classification map of urban functional zones can provide valuable support for urban planning, management and decision-making. Compared with the traditional land cover and land use, urban functional zones contain more complex spatial structure and semantic information, which brings great challenges to their discrimination. In this study, a new framework that can integrate very high-resolution (VHR) satellite images and POI data is proposed for functional zone recognition. It uses the street blocks enclosed by the road network as the basic functional units. Further, the deep learning model SE-ResNet and the natural language processing model word2vec are used to learn the image spatial features and POI semantic features of street blocks respectively. Finally, the two features are fused and input into SVM classifier to realize the recognition of functional units. The classification experiment of urban functional zones in Futian District, Shenzhen, China verified the effectiveness of our proposed method. The overall accuracy of classification using fused features is 86.83%, which is 5.83% and 12.83% higher than that using only image features and semantic features, respectively. For each functional category, the classification accuracy has also been improved to a certain extent.