A Study On Intelligent Building Polygonal Contour Extraction From High Resolution Remote Sensing Image

Building information is essential for numerous applications,encompassing urban planning,population density estimation,change detection,energy supply,and disaster management.High-resolution remote sensing images can provide abundant geospatial information and furnish a reliable data foundation for high-precision extraction of polygonal building outlines.The advancement of deep learning-based methods has significantly boosted the performance of automatic extraction of buildings.However,automatic extraction of regular vectorized polygons from high-resolution remote sensing images remains an arduous task owing to the diversity of building structures and variations in imaging conditions.This paper proposes a deep learning-based technical roadmap to delineate regular polygonal building contours from highresolution remote sensing images.The specific research contents include the following points:(1)Research on Polygonal Building Contour Extraction Based on Semantic Segmentation and Regularization Post-Processing.Most current building extraction research focuses solely on pixel-level semantic segmentation.The proposed approach combines deep learning methods with artificial experience design methods to enable the intelligent extraction of building rule vector polygons that satisfy the needs of practical applications,based on building semantic segmentation.The method comprises two main components: semantic segmentation of buildings and a regularization algorithm grounded in empirical design.First,to address the issues of weak aggregation of image feature information in semantic segmentation networks and poor robustness of object extraction across scales,a multi-scale feature aggregation Fully Convolutional Networks(MA-FCN)is developed for building semantic segmentation.Additionally,two simple and effective post-processing strategies are introduced to achieve high-precision building semantic segmentation map extraction.Second,a building polygon regularization algorithm based on empirical design is proposed,which adjusts the contour polygons based on the main building direction to obtain regular building vector polygons.Experimental results on multiple datasets demonstrate MA-FCN’s superior performance in building semantic segmentation.The regularization algorithm is robust to challenges such as different building types,images of varying resolutions,and semantic segmentation outputs of different quality.After regularization and adjustment,the building polygons can achieve the level of manual labeling by operators.(2)Research on Polygonal Building Contour Extraction Based on Contourbased method and Regularization.The methods proposed in this study aim to extract building vector polygons directly using contour-based techniques,overcoming the limitations of "semantic segmentation followed by vectorization" approaches.The study consists of two stages: building polygon extraction based on two-scale graph convolutional networks(TS-GCN)and building polygon extraction based on concentric loop convolution network(CLP-CNN).In the first stage,GCNs are used to extract building polygons.A two-scale GCN is introduced to improve contour extraction accuracy.First,individual building bounding boxes are obtained,and contour initialization is performed based on specific rules.Then,TS-GCNs are used to extract high-precision building polygons.In the second stage,CLP-CNN is proposed as an improvement,distinguishing itself from methods using only single-ring edge information for contour localization.CLP-CNN can fully capture instance edges and internal information for precise edge localization.The network first predicts initial building contours in polar coordinates,and then optimizes the contour coordinates using concentric loop convolutional structures.A vertex refinement module removes redundant points far from building corners.Finally,a regularization algorithm is applied to adjust the vector polygons obtained from the two methods to obtain regular building polygons.(3)Research on Polygonal Building Contour Extraction Based on end-to-end contour-based method.This study proposes an end-to-end contour-based method named Build Mapper to extract regular vector polygons of buildings.Build Mapper achieves end-to-end learning from remote sensing images to regular building vector polygons without relying on additional rule-based post-processing,making it applicable to a wider range of building scenarios.The Build Mapper network has a concise structure,high extraction accuracy,and fast inference speed.It consisting of two main modules: 1)a contour adaptive initialization module,which replaces the pre-defined edge initialization with an adaptive approach to generate initial building edges based on the image features,and 2)a contour optimization module simultaneously performs contour coordinate adjustment and redundant vertices removal.It utilizes a lightweight feature encoding structure to capture vertex adjustment information and employs a vertex classification head to remove redundant vertices,resulting in a polygon composed only of building corners.This eliminates the need for complex empirical post-processing.This method can also be used for semi-automatic annotation of building polygons,where the annotator marks the building center point,and the algorithm automatically outputs the regular vector polygon of the current building instance.(4)Research on Polygonal Building Contour extraction method from line detection to polygon reconstruction.The study proposes a novel bottom-up approach,Line2 Poly,to extract buildings vector polygons using building feature lines as the foundational elements.Line2 Poly also considers extracting internal building feature lines,making it suitable for a wider range of applications.Line2 Poly consists of two main parts: building feature line extraction and polygon topology reconstruction.First,a convolutional neural network(CNN)is used to initialize candidate building feature lines,which are then used as positional indices to initialize the input of a transformerbased module for accurate extraction of high-precision building feature lines.Subsequently,a topology reconstruction module based on feature line relationships is employed to determine the adjacency relationships between discrete building feature lines,thereby reconstructing the building polygons.Extensive experiments on multiple open-source datasets demonstrate that Line2 Poly achieves state-of-the-art performance in building feature line extraction and polygon extraction tasks.Notably,Line2 Poly achieves breakthroughs in automatically extracting building vector polygons with human-drawn accuracy on multiple datasets without any manual intervention.