| An accurate classification map of urban function zones plays a crucial role in supporting urban planning and development.Currently,the identification of urban function zones primarily relies on remote sensing images,social sensing data,or the fusion of multiple data sources.Very highresolution remote sensing images possess optical characteristics that can reflect spatial information such as the shape,structure,and spatial arrangement of the city.However,such images always lack semantic information regarding urban functions.On the other hand,social sensing data can describe the semantic features of urban function zones effectively but is deficient in spatial information and has limited perception of hybrid urban functions.By fusing features from multiple data sources,the methods used to identify urban function zones can overcome the shortcomings of relying solely on a single data source,and optimize the classification effect of urban function classify method.Urban planning and human activities genertate complex spatial topological relationships between adjacent urban function zones,and these relationships can be further deepened through spatial analysis methods to understand urban function zones better.In this study,a new framework which can intergate very high-resolution remote sensing images and open social data(including POI and OSM road networks)conbine with spatial topological relationships is proposed for urban function zones recognition.The main work includes the following two parts:(1)To deal with the problem of poor classification performance of methods only using feature from single data source,a new method for identifying urban function zones by fusing visual features of very high-resolution remote sensing images and semantic features of POI data proposed to achieve fine-scale classification of urban function zones.Based on Open Street Map data,GF-2 satellite images,and POI data in Shenzhen,Guangzhou,China,firstly constructs the road network of Shenzhen and divided into street-block as the basic function units.Then,the convolutional neural network SERes Net50 is used to extract visual features,and the natural language processing model Word2 Vec is used to extract semantic features.Finally,the two features are fused to achieve end-to-end classification recognition of urban function zones.Experimental results show that the overall accuracy of classification urban function zones based on fusion features is 85%,which is better than the accuracy using single data features(72% for visual feature and 77% for semantic feature).Verified this method can ameliorates the problems of using single features in urban function zones classification.(2)To make full use of spatial relationships between urban function zones,a framework incorporating spatial relationship features is proposed.The complex spatial topological relationships between urban function zones can be described by a graph.First,we construct a graph structure which the nodes of this graph are basic street blocks,and the initial features of each node are fusion features.Then a graph convolutional neural network is utilized to learn the relationships between adjacent function zones.The spatial relationship features,visual features,and semantic features are fused to achieve end-to-end classification of urban function zones.The final experimental results show that compared to the method of using fustion visual and semantic features,this method can improve accuracy in the whole study area and all function categories,the overall accuracy is 88%,which have a 3% increase than using fusion visual and sementic features.The above results show that fusing visual features,semantic features,and spatial relationship features can achieve the closest results to the actual classification of urban function areas,and fused features can more accurately depict the characteristics of each urban function zone,providing more precise,rapid,and comprehensive support for urban planning and management. |
PDF Full Text Request