Research On Building Extraction From High-resolution Remote Sensing Images Based On Deep Learning

Buildings are an important and active element of urban development,and are the focus of attention in the fields of land resource management,urban planning,digital cities,and 3D modeling.Efficient and accurate extraction of building information has important theoretical and practical significance for these fields.Due to the high information redundancy and complex and diverse features of high-resolution remote sensing images,how to design better automatic building extraction algorithms has become a challenging and widely concerned research direction.Deep learning methods have shown excellent performance in the field of building extraction research.However,the existing deep learning methods are difficult to extract large-scale distribution of complete buildings due to the irregular shape and large differences in the appearance of buildings.At the same time,because high-altitude imaging is prone to complex occlusions,shadows and other problems,it is also very easy to lose the detailed information of small buildings during extraction.In order to improve the accuracy of building extraction,this paper optimizes the existing network model from two aspects of multi-scale information and global context information.Set to carry out experiments,carry out verification and analysis.The main research contents and results of this paper are as follows:(1)Based on the multi-scale information of high-resolution remote sensing images,a new building extraction method is proposed.First,the spatial pyramid module is added to the feature extraction part of the U-Net network to obtain feature maps of different scales,and then the SENet module is selected to add an attention mechanism to the channel dimension to improve the ability of the network model to represent features.information is passed to the decoder.The accuracy rate,F1-score and Io U of the improved network structure in this paper reached 98.61%,90.33% and95.48% respectively on the WHU dataset,and 95.33% and 72.05% on the Massachusetts dataset.,86.77%,all ranked first among the models participating in the comparison,which verifies the effectiveness of the method,and realizes the full fusion of multi-scale information of remote sensing images and the improvement of building extraction accuracy.(2)A building extraction method based on high-resolution remote sensing image context information is proposed.Since the convolutional neural network will lose a lot of local information through repeated convolution and pooling operations,this paper proposes a new encoder-decoder structure building extraction network based on this idea,inserting non-local modules into the network to introduce The global information improves the extraction accuracy of the model,and the residual network module is used as a feature extractor to ensure that the network can fully obtain the pixel context information.The experimental results show that the accuracy,F1-score and Io U of the method on the WHU dataset reach 98.61%,90.33% and 95.48%,respectively,while on the Massachusetts dataset it is 97.5%,67.35%,89.24%,which is similar to Compared with the unmodified U-Net network,it can also achieve 1.06%,1.44%,and 2.31% accuracy improvements on the high-resolution and well-labeled WHU dataset,respectively.On the lower-resolution Massachusetts dataset,It is at least increased by 3.7%,2.07%,and 3.63%,respectively,which fully shows that the improved network model enhances the ability to obtain contextual information from high-resolution remote sensing images.Figure[28] Table[6] Reference[79]...