High-resolution Remote Sensing Building Extraction In Complex Urban Scenes

Although the land area of cities is only 2% of the earth's surface,more than half of the world's population lives in urban areas.The rapid development of cities has caused the number of urban buildings to rise rapidly.Therefore,it is of great significance to accurately extract urban buildings from high-resolution remote sensing images.High-resolution remote sensing building extraction in complex urban scenes has the problem of being easily blocked by shadows or vegetation and multi-scale effects,which brings huge challenges to refined extraction.This paper uses advanced deep convolutional neural networks,combined with dual attention module and composite loss function,to improve the performance of the U-Net model for high-resolution remote sensing building extraction in complex urban scenes,and use it in multi-source remote sensing datasets and various good results have been achieved in complex urban scenes.The main contents of this article are as follows:(1)Analyze the remote sensing characteristics of urban buildings.Urban buildings are quite different from ground objects such as shadows and vegetation in terms of spectrum,texture,and geometry.There is a certain difference between the multi-scale building features obtained by using fixed receptive fields in deep convolutional neural networks.In the face of shadow or vegetation occlusion and multi-scale effects,although conventional deep learning models have achieved certain advantages,the extraction performance of building features needs to be enhanced.The dual attention module can self-adjust to acquire global and long-channel building features based on the similarity between building features,thereby enhancing the building features under shadow or vegetation occlusion,and multi-scale effects.(2)Constructed a high-resolution remote sensing urban building extraction model.Traditional urban building extraction methods can only extract shallow features,and methods based on deep learning are insufficient for the feature extraction of buildings in complex urban scenes.In response to the above problems,this article improves the U-Net network.First,a dual attention module is embedded at the end of the encoder of the standard U-Net network to capture global building information and long-channel building information,and enhance the building characteristics.Then the loss function is improved,and the Lovász loss function is added to the cross-entropy loss function.The composite loss function formed improves the model's ability to constrain the building extraction results and further enhances the robustness of the model.(3)The applicability of the method in this paper is verified in complex urban scenarios.FCN,U-Net,Seg Net,Deep Lab V3+,and the method in this paper are verified on the Inria Aerial Image Labeling Dataset,Massachusetts buildings dataset,and World View-2 data respectively,and the results under shadow or vegetation occlusion,and multi-scale effects are selected for analysis.The study found that the method in this paper has the best extraction performance for buildings in complex urban scenes,and can effectively solve the problem of missing or incomplete extraction of buildings in the case of shadows or vegetation occlusion,and multi-scale effects.The F1-score on the Inria Aerial Image Labeling Dataset reached 85.54%,and the F1-score on the Massachusetts buildings dataset reached 83.78%,and relatively good results were also achieved on the World View-2 data.