Research On Building Extraction From Optical Satellite Remote Sensing Images

With the acceleration of global urbanisation and population growth,buildings have become one of the most widely ground objects.Using semantic segmentation in building extraction in remote sensing images can quickly obtain the location and texture information of buildings,thus enhancing the capability of civil applications such as urban planning,disaster monitoring and intelligent navigation,and promoting the development of information technology construction.In addition,with the development of satellite remote sensing technology,satellite remote sensing images have the characteristics of wide range,high realtime,relatively inexpensive,and are not restricted by national boundaries,airspace and terrain;not only that,compared with other data types,optical images are easy to obtain,easy to process,easy to understand,and have the advantages of high resolution and rich information such as texture and colour.Therefore,the extraction of buildings from optical satellite remote sensing images has become a very important research direction in the field of remote sensing image analysis.The main challenges in extracting building features based on optical satellite remote sensing images are as follows:(1)Complex background of optical satellite remote sensing images;(2)Inaccurate positioning of building edges;(3)Inadequate use of building information.In addition,most current networks are based on the assumption that both training and test data belong to independent and homogeneous distributions,i.e.the same dataset,which undoubtedly ignores the cross-domain problem,As a result,there is a significant degradation in model performance when testing out-of-distribution datasets in the face of real-world application situations.This paper investigates the following elements in response to the above problems:1.Under the background condition of same spatial scale domain data set,starting from the characteristics of images and buildings themselves,This paper studies how to optimize the results of building extraction and proposes SD-Net(Standard and Dilated Convolution Network)network.Firstly,aiming at the complex background of remote sensing image,DICE + Loss is proposed to enhance the model ’s attention to the True Positive part of building extraction and avoid the interference of complex background.Secondly,aiming at the problem of inaccurate edge positioning,the boundary loss function and DICE + Loss are added to form a weighted loss function to optimize the extraction of building boundaries.Finally,aiming at the problem of insufficient utilization of building information,a feature extraction module of convolution fusion is proposed,focusing on the detailed texture and spatial structure features of buildings.The experimental results show that SD-Net can effectively solve the above problems,and the performance is better than other comparison algorithms.2.Under the background conditions of cross spatial scale domain data sets,For optical satellite remote sensing images,data sets captured in different regions,different resolutions,and different periods generally have differences in appearance features such as color,texture,and resolution,from the above feature differences,study how to solve cross-domain problems.Firstly,a data augmentation method based on random convolution is designed to generate new domains to introduce more prior knowledge into the model and improve the performance of cross-domain extraction.Secondly,a new residual unit Basicblock A is designed,which reasonably uses instance normalization to improve the residual unit to eliminate the feature difference in appearance,and then realizes feature alignment.According to the two parts of the design,the SD-Net + network is proposed.The experimental results show that SD-Net + can effectively solve cross-domain problems.