Recognition Method Of Building Rooftops In High Spatial Resolution Remote Sensing Imagery

As an important artificial feature(geographic elements),the spatial distribution information of buildings reflects the evolution layout of human social and economic development.The scientific and accurate measurement of the spatial distribution and dynamic changes of building rooftops lays a solid foundation for the establishment of building survey and update mechanism,and the promotion of regular building monitoring,which has important theoretical research significance and practical application value.The complexity of high spatial resolution remote sensing image data makes the traditional building rooftop recognition methods have the limitations of low automation efficiency and recognition accuracy,and the deep neural network using computer vision solution idea has achieved better practical application effect,but there are still many problems to be solved,especially the following two problems are the core:(1)The spatial-temporal genealogical features of building rooftop challenge the construction of deep learning sample libraries for building rooftops;(2)The geometric,spectral heterogeneity and multi-scale features of building rooftops challenge the accuracy and generalization ability of deep learning recognition models.This paper is oriented to the needs of industries such as the recognition and change monitoring of building rooftops by high resolution remote sensing,and addresses the problems of recognition errors and omissions caused by the multi-scale symbiosis and spectral heterogeneity of building rooftops in high resolution remote sensing images.On the basis of the pedigree feature representation of the target sample library of building rooftop images and the calculation of multi-scale granularity feature information,the multi-scale recognition of building rooftops in high resolution remote sensing images is accomplished by relying on techniques such as graph theory and deep learning,and the multi-scale recognition method and vector contour polygon extraction technology system of building rooftops are constructed.The research contents and findings of this paper are as follows.(1)Image pre-processing and feature sample library construction for building rooftop pedigree features.In this study,image fusion,image cropping,and image enhancement based on linear stretching and spatial domain filtering are used to purify,enhance,compress and normalize the high-resolution remote sensing images to obtain image datasets that satisfy the deep neural network model.Then,we construct a pedigree deep learning feature sample library adapted to the target of building rooftops for the geometric and spectral visual feature differences and attention strength of different building rooftops on the high-resolution remote sensing images,and lay the data foundation for the training and prediction of multi-scale building rooftop depth recognition network models.(2)Multi-scale recognition of building rooftops based on a multi-scale deep neural network model.Supported by theoretical techniques such as remote sensing information graph and deep learning,this study designs and develops a stacked multi-scale granularity instance segmentation network model MS-CNN with multi-scale granularity feature extraction algorithm as a tool for calculating the scale features of high-resolution remote sensing building rooftops.The model training of MS-CNN is carried out based on the spatio-temporal pedigree sample library of building rooftops,and the instance segmentation of building rooftop target image primitives is carried out to complete the multi-scale symbiotic recognition from remote sensing images,and the effectiveness and robustness of the model for multi-scale target recognition are verified through experiments.(3)Empirical evidence of building rooftop recognition in Beijing Municipal Administrative Center.This study uses the multi-scale building rooftop recognition model MS-CNN constructed by iterative training,and carries out the empirical demonstration of building rooftop pedigree multi-scale recognition in Beijing Municipal Administrative Center area with the reference base of high-resolution remote sensing image background data,and completes the production of building rooftop thematic data in Beijing Municipal Administrative Center area by using vector polygon contour extraction technology.It is focused on the establishment of a multi-scale recognition method and polygon contour extraction technology system for building rooftops and strive to explore the way for the application of high-resolution remote sensing image building rooftop recognition engineering.