Multi-Scale Urban Function Recognition Based On Multi-Source Spatiotemporal Data Fusion

In China,rapid urbanization process has profoundly influenced economic development and social progress,also puts forward severe challenges to urban planning,traffic congestion,resource allocation etc.,thus making it significant to deeply understand urban spatial layout.With the emergence of massive spatiotemporal data,remote sensing images,social media data and other big data have been important data sources to finely characterize the urban environment.Also,it has triggered a series of new urban research paradigms,e.g.urban computing,social sensing and urban informatics.Existing multisource spatiotemporal data fusion models mostly ignore the interaction among features and its scale effect.Meanwhile,urban function recognition was mainly concentrated on a single scale,which makes it difficult to form a systematic perception of urban spatial layout.Thus,it is vital to effectively fuse multi-source spatiotemporal data to accurately and comprehensively sense urban function.To this end,based on high spatial resolution(HSR)remote sensing image,point of interest,and Tencent user density data,this study establishes a coupling framework between multi-source spatiotemporal data fusion and multi-scale urban function recognition,including the following three parts:(1)Inspired by the correspondence of multi-source spatiotemporal data,this study utilized the "bag of words" model and probabilistic topic model to extract the physical and social semantics from multi-source spatiotemporal data,and then canonical correlation analysis and kernel canonical correlation analysis were used to generate the linear and nonlinear cross-correlation for land patches.This model makes up for the demerits of one-source spatiotemporal data,e.g.bias and uncertainty,thereby realizing the effective fusion of multi-source spatiotemporal data.(2)Inspired by the multi-view descriptions of physical semantics and social semantics,this study combined semantic information and cross-correlation to carry out multisource feature fusion,and then the CC-FLU model was built to classify urban function after evaluating features’ importance.Through the comparative analysis of nine experiments in three hierarchies,the CC-FLU model obtained a better classification accuracy than some typical models,e.g.scene classification model and social sensing model,and had better performance in the recognition of various land functions.(3)Inspired by the scale issue of multi-source spatiotemporal data,this study proposed a multi-scale spatiotemporal data fusion framework based on a dynamic semantic weight model.The appropriate feature combination was selected by exploring the scale effect of fusing physical and social semantics,and then multi-scale urban function was recognized.The results reveal that the physical and social semantics emphasize the low and high scales,respectively.Meanwhile,in the transformation from the low to the high scales,land functions gradually evolve from physical attributes to social attributes.In general,this study provides a new perspective to explore the fusion mechanism of multi-source spatiotemporal big data.Based on the correspondence problem and the scale problem of multi-source spatiotemporal data fusion,this study derives two effective fusion models to carry out high-precision urban function recognition.Meanwhile,the result can provide fine functional drawings for urban planners and managers to accurately and comprehensively understand urban spatial layout,which is important for urban renewal,resource allocation and so on.