| In recent years,with the slogans of "solving the ’three rural’ problems","revitalizing the countryside" and "building a livable,workable and beautiful countryside " and other slogans,the development of rural economy has become an inseparable topic in China’s modernization.To a certain extent,buildings reflect the social and economic development,and are the most frequently changing and most valuable type of features in the geographic information database.At present,the traditional remote sensing image building extraction requires manual design features with low accuracy ceiling;the existing building semantic segmentation dataset has single data type and low resolution,the semantic segmentation method cannot fully exploit its data performance,and the extraction result boundary is irregular and cannot be directly used for mapping production.In addition,the style of buildings around the world varies,and the texture information of buildings varies greatly from place to place,but the style of buildings in a small area is similar.For these reasons,this thesis designs a rural building extraction and contour fitting model based on U-Net++ for a village building in Xuzhou to improve the accuracy and efficiency of rural building extraction,and the semantic segmentation generalizability problem can be avoided by taking a small range of buildings as the research object.The main works of this thesis are as follows:(1)The semantic segmentation datasets of color buildings can only use texture information to extract buildings but not elevation information.In this thesis,based on the UAV tilt photography technology,we obtain DOM and DSM images with 3 cm resolution in the study area,and use the image fusion technology to obtain DOM and DSM fused images,and label the buildings in the three images respectively to make a semantic segmentation dataset of buildings with small samples supported by multi-domain features.(2)Comprehensive comparison of various images semantic segmentation algorithms,U-Net++ is selected for rural buildings extraction.To further improve the performance of the network,a variety of classical strategies are selected to optimize the network in this thesis.Through multiple sets of experiments,the optimal data set type and the most reasonable training set share are determined,taking into account the segmentation accuracy and the cost of data set annotation.The experimental results show that the Io U value of fused image dataset with about 12% as the training set is similar to that of DOM image dataset with about 24% as the training set and DSM image dataset with about 33% as the training set,i.e.,it is the most economical to use fused images dataset for buildings semantic segmentation.(3)Given the presence of noise inside the raster map extracted by semantic segmentation,the complete buildings are cut into multiple pieces.In this thesis,the raster map is processed by images binarization,and the interior of the building is denoised using expansion and erosion algorithm,and then the result is stitched and vectorized.(4)To address the problems of smoothed representation of key nodes and node redundancy in the vectorized building contours.This tthesis proposes a contour fitting algorithm suitable for right-angle buildings.The fitting algorithm first classifies the edges of the vector map with Otsu’s algorithm,then removes the discrete points with large deviations with the triple standard deviation law,then fits a straight line with the least squares method,then determines the position relationship between two straight lines with the slope of the fitted line,and finally connects the vertices in turn with the intersection of the reconstructed line segments as the vertices of the vector map to finally complete the contour fitting of rural buildings.The contours of the rural buildings are finally fitted.The redundant nodes are removed,the boundary is leveled,and the key nodes are reconstructed and fit closely to the buildings.This thesis contains 46 figures;9 tables;87 references. |
PDF Full Text Request