| With the continuous acceleration of urbanization,the idea of sustainable construction has gradually become the main theme.Among them,buildings are the important symbols of urban development.Automatic extraction of buildings plays a crucial role in urban planning,national defense construction,land use change monitoring,disaster monitoring and early warning,and has an important reference value for smart city construction.With the resolution of remote sensing images reaching sub-meter level,more and more fine structure,texture and spectral information of buildings can be obtained,and high-resolution aerial images become one of the important data sources for building extraction.However,the large intra-class variance and small inter-class variance of high-resolution aerial images lead to difficulties in designing classification features manually.And also,the complexity of surface buildings makes the accurate and automatic building extraction a challenging task.As deep learning techniques show powerful advantages in the field of natural image processing,more and more researchers try to apply deep learning techniques to the field of remote sensing image analysis.The powerful feature extraction ability of deep learning techniques can obtain the abstract representation of images,which helps to obtain deeper semantic information of buildings and get more accurate building extraction results.However,the current building extraction methods based on deep learning still have the following problems:first,the pixel-by-pixel prediction of traditional convolutional neural networks does not make enough use of the relationship between pixels,which leads to problems such as discontinuous segmentation results in the existing models for extracting regular buildings;second,the lack of global dependencies in traditional convolutional neural networks leads to problems such as boundary blurring or boundary noise in the existing models for extracting irregular buildings;third,the adaptive aggregation capability of traditional convolutional neural networks for features is insufficient,which leads to the lack of completeness of existing models when identifying small buildings.For the above-mentioned problems,this thesis proposes a corresponding solution based on the ideological guidance of integrating global and local information:(1)To address the problems of discontinuous segmentation results of existing models for extracting regular buildings,this thesis proposes a generative adversarial segmentation network(ASGASN)that incorporates atrous spatial pyramidal pooling and skip connections.The core of the method is based on the adversarial training strategy to strengthen the relationship between the focus pixels and improve the continuity of the building extraction results.At the same time,the method introduces depth-separable convolution and global convolution to improve the classification and localization accuracy of the model,and uses the atrous spatial pyramid pooling module to enhance the model’s ability to perceive buildings at different scales.Based on the above strategies,this method possesses more accurate regular building extraction capability than traditional neural networks.(2)To address the problems of boundary blurring or boundary noise when extracting irregular buildings by existing models,this thesis proposes a building graph convolution network(BGC-Net)that combines a deep full convolutional network and graph convolution.The core of the method is to propose a dual graph convolution module,which can model the contextual information in space and channels to obtain global dependencies.At the same time,the method constructs the atrous attention pyramid module,which embeds the attention mechanism into the pyramid structure to obtain the multi-scale features of buildings more accurately.Due to the guidance of global dependency relations and the acquisition of multi-scale features,this method can obtain more clear and accurate irregular building boundaries.(3)To address the problems of insufficient completeness of existing models to identify small buildings,this thesis proposes a multi-scale building segmentation network(SCA-Net)incorporating dual attention mechanisms.The core of the method is to embed spatial and channel attention mechanism modules that adaptively aggregate building features.At the same time,the method proposes a dense connection feature pyramid module,which enables dense spatial sampling under large receptive fields to obtain a larger range of multi-scale features of buildings.Compared with traditional neural networks,this method can adaptively aggregate building features and obtains more complete small building extraction results.In this thesis,the proposed three methods are experimentally validated based on the WHU Building Dataset and the China Typical City Building Dataset.The results show that the overall accuracy of the three methods proposed in this thesis is above 0.925 on both datasets,and all of them can effectively solve the corresponding problems.Among them,the ASGASN-based building extraction method has the highest accuracy of regular buildings extraction on both datasets;the BGC-Net-based building extraction method extracts irregular buildings with more regular and clearer edges;the SCA-Net-based building extraction method can obtain more complete small buildings.This thesis proposes corresponding solutions to the problems existing in building extraction,explores a new way for the building extraction method based on contextual semantics,and provides a theoretical reference for the intelligent interpretation of geomatics in the era of remote sensing big data. |
