Optimization Method For Building Rooftops Recognition With Collaborative Tupu Theory In High Spatial Resolution Remote Sensing Imagery

With the rapid development of remote sensing technology,the amount of image data is increasing and the data types are becoming more and more abundant.Buildings are one of the most important geographic elements in remote sensing of resources and environment,and the automated extraction of building roof information is of great significance in remote sensing of land and resources and other related application fields.Building roofs in high spatial resolution remote sensing images have significant geometric and spectral features,and the information of building roofs can be obtained quickly and accurately in a large area by using remote sensing technology.Deep convolutional neural network,with its excellent deep feature extraction capability,overcomes the problem that traditional remote sensing image target recognition methods are difficult to effectively obtain the complete physical image primitives,and has achieved good application results in intelligent recognition of building roofs,but there are still some key scientific problems that have not been effectively solved:(1)the edge transition area formed by mixed pixels on both sides of the roof boundary in the high-resolution image,which leads to the inaccurate segmentation results at the boundary by the deep learning network model,(2)the width direction of the roof raster boundary is composed of multiple pixels,which leads to the extracted polygon contour lines of the building roof segmentation in the ‘raster vector’conversion process usually has ‘jagged’,(3)the geometric spectral features of roofs are easily confused with side elevations,trees and shadows,and so on,and the general method is prone to the problem of error and omission detection.In this paper,we have researched and analyzed the optimization method of building roof profile for the problem of inaccurate segmentation at the boundary caused by the mixed pixel transition area of building roof edge in remote sensing image,and the main research work is as follows:(1)Construction of a mask regularization mechanism model for the transition zone of building roof edgesIn this study,we propose an interpretable statistical model-guided generative optimization model MSBR-GNet for the boundary spatial domain and spectral domain,and add the statistical model-guided boundary loss function combining spatial domain and spectral domain into the generator loss calculation of MSBR-GNet to precisely constrain the regularized generation of building roof contour masks by the interpretable mechanism.The experiments show that MSBR-GNet can achieve a more regular extraction of building roof contours,and the Precision values of 0.9275,0.9228 and 0.8779 in the public datasets of Inria,WHU and Massachusetts,respectively,which can guarantee the accuracy of building extraction and achieve the best results in the boundary morphology evaluation index.(2)Construction of a sample library(set)of building roof genealogy mapsBased on the public building datasets of WHU,Inria,Massachusetts and the own datasets of remote sensing images of representative cities in China,this study proposes a genealogical sample library(set)construction method based on the three laws of geography(correlation,heterogeneity and similarity)and the map theory,and accordingly implements the genealogical sample library(set)of building roofs.BUCEA2.0(Open source).The feasibility of BUCEA2.0is verified through the ablation experiment of the genealogical sample set,and the migration of the MSBR-GNet model with both geographical coupling and global universality is realized from the perspective of the interpretability of the sample base(set).(3)Empirical evidence of building rooftop recognition contour optimization in Beijing Municipal Administrative CenterThis empirical study uses BUCEA2.0,a sample library(set)of building rooftop pedigree profiles,to train the constructed MSBR-GNet,a generative optimization model of building rooftop boundaries guided by interpretable statistical models in the boundary space domain and spectral domain,to carry out the empirical demonstration of MSBR-GNet in the Beijing Municipal Administrative Center area,and to complete the 2022 Beijing Municipal Administrative Center building.The production of thematic data on roofs of buildings in the2022 Beijing Municipal Administrative Center was completed.