| Buildings are important places for human activities,and their information is an important credential for understanding urban development.Remote sensing satellites can obtain large-scale,high-real-time surface observation data.Especially,high-resolution remote sensing images can clearly display the elements of urban spatial features,which are helpful to grasp the current scale and layout of urban buildings.The accurate extraction of buildings from high-resolution remote sensing images has substantial application value in building census,law enforcement against illegal buildings,urban development planning,high-precision map update,and earthquake disaster assessment.In the high-resolution remote sensing images of urban scenes,the spectrum,shape,texture,and scale of buildings vary greatly,and there are problems such as interference from similar ground objects,dense distribution of buildings,and blurred boundary lines.The above situation limits the application of various methods in building extraction from high-resolution remote sensing images.This paper aims to improve the technical level of high-resolution remote sensing image building extraction technology,summarizes the research status of traditional methods and deep learning methods for high-resolution remote sensing image building extraction,and analyzes the limitations of each method.In terms of specific technical implementation,this paper studies three methods: traditional supervised classification,deep learning semantic segmentation and instance segmentation.The research results are as follows:(1)On the basis of traditional supervised classification,A multi-level optimization extraction method of urban buildings from high-resolution remote sensing images is proposed.Features are the decisive factor affecting the performance of building extraction.High-resolution remote sensing images have large differences in buildings in urban scenes,and there are similar ground objects interference.Supervised classification automatically learns various ground object features from samples,which is conducive to extracting various types of buildings.This paper uses the traditional maximum likelihood classification algorithm to extract buildings,and then aiming at the problem of false detection of extraction results and "salt and pepper effect",based on the methods of morphological building index and object-oriented information extraction,as well as the feature performance of each false detected ground object in the image,constructs multi-level optimization rules to extract more accurate building.The experimental results show that this method can better balance the integrity and correctness of building extraction,and effectively improve the accuracy of building extraction.(2)A method of building extraction from remote sensing images based on a multi-scale encoder-decoder network is proposed.The buildings in the image have large size differences,dense distribution and fuzzy boundary lines.The semantic segmentation network needs to take into account the global semantic features and local detail features of buildings of different sizes,so as to avoid problems such as missing mention,incomplete extraction,false mention,and inaccurate boundaries.In this paper,the encoding and decoding architecture of UNet network is adopted,and the convolution blocks of VGG16 are stacked in series to form the encoder backbone network,which is used to extract deep semantic features,and the fifth convolution block is changed into multi-scale module to improve the ability of the network to extract deep scale features;Then the connection between encoder and decoder is redesigned to make full use of the deep and shallow scale information in the encoder;Finally,an additional constraint is added to the decoder to improve the ability of network training and strengthen multi-scale feature representation.The experimental results show that the method can effectively improve the problems of incomplete building extraction and inaccurate boundaries.On the WHU aviation dataset and the Massachusetts dataset,the Io U reached 89.45% and 70.02%,respectively.On the Guilin GF2 dataset,the Io U was 3.53% and 2.72% higher than that of Deep Lab V3+ and UNet,respectively.(3)A building extraction method based on model fusion and image segmentation is proposed.In complex urban scenes,the building patches extracted by the pixel-level semantic segmentation network are incomplete.This paper takes the classic instance segmentation model Blend Mask in deep learning as an example to carry out the research on building object-level extraction.First,we explore the effect of the number of Res Net layers on the extraction of buildings by Blend Mask.In view of the advantages of deep and shallow network models,this paper adopts the idea of model fusion,introduces image segmentation,and adopts the idea of majority voting for each superpixel to optimize the building patches.The experimental results show that the method achieves the target-level extraction of buildings through model fusion and image segmentation,and the extraction of single buildings is relatively complete,with less patch adhesion,and further improves the accuracy of building extraction. |
