Abstract: Detailed classification of building function provides scientific support for urban planning and digital city modeling.However,existing researches primarily focus on identifying the geometric attributes of buildings based on high-resolution remote sensing images,paying less attention to their functional attributes,and largely ignoring the increasingly popular mixed-use buildings.To address this,this paper introduces multi-source geospatial data,including building vector geometry data,building vector environmental data,and POI data,to compensate for the limitations of optical,nighttime light,and SAR remote sensing images in depicting building uses.A multi-modal data fusion model,MultiMixNet,is proposed to achieve simultaneous identification of both single and mixed-use buildings.Experimental results in urban areas of Shanghai,Hangzhou,and Xi'an show that the MultiMixNet method achieves an average Micro F1 score of 76.63% and an average Macro F1 score of 71.47% across the three study areas.The identification accuracy for mixed-use buildings generally exceeds 60%.The nighttime light data significantly improving the classification accuracy of commercial buildings by reflecting nighttime activity characteristics,while building vector environmental data provides crucial support for distinguishing between residential and mixed-use commercial-residential buildings.To address the issue of mixed-use buildings being far fewer in number than single-use buildings,this paper designs a building buffer zone extraction module based on the object detection framework to enhance the sample proportion of minority categories,effectively alleviating data imbalance and improving the classification accuracy of mixed-use buildings.

Key words: building function identification, mixed-use buildings, building environmental features, deep learning