Building Change Detection Based On Multi-branch Structure Model Of Deep Learning Method

As one of the important ground features,buildings are closely related to human life.The change detection is of great significance in the land resources planning task of China’s land and resources department.However,the following problems still exist in the task of building change detection: there are many background information and redundant feature information;Moreover,the building itself is complex and diverse,and has the same spectrum characteristics with other impervious surfaces,which will lead to the problem of "same spectrum change";Affected by the shooting angle of satellite sensors,buildings have different forms in remote sensing images;At the same time,the image position deviation in different periods is also easy to cause false changes in the detection process.When traditional methods detect building changes,for a single scene,they can not achieve efficient feature learning for building changes,which makes the recognition results have some problems,such as low integrity,inaccurate boundary and so on.Based on the problems existing in the above change detection tasks,this paper designs a two branch network deep learning model,verifies the effectiveness of the method and application,realizes the building change detection of the main streets of Binhu New Area in Hefei,and makes a temporal and spatial analysis,in order to provide a theoretical basis for the follow-up government department resource planning.The main achievements are as follows:(1)According to the image characteristics of building changes on domestic highresolution remote sensing satellite,the construction rules of sample database are formulated,and three main types of buildings are established: demolition,new construction and reconstruction.We have completed the preprocessing of aviation data sets and domestic high score series satellite data sets,and completed the production of sample sets.(2)Aiming at the problems of low recognition accuracy and incomplete extraction boundary in the task of building change detection.This paper designs a deep learning model of double branch structure,and uses dense blocks to connect the whole network in series,so as to reduce the amount of network parameters and over fitting in the training process.By using the feature enhancement module,and using the main branch to learn the change features,the auxiliary branch guides the building feature learning,so as to realize the effective feature learning in the specific semantic environment of building change.In order to reduce the loss of shallow information in the down sampling process of deep learning network,this paper adopts jump connection and global information enhancement mechanism in the process of network modeling to retain the detailed information of building changes.(3)The effectiveness of the proposed network is verified on the aeronautical data set,domestic high-resolution identical sensor image pairs and different sensor image pairs.Combined with the pixel based accuracy evaluation method and the object-based accuracy evaluation method,the network designed in this paper shows good results.Compared with densinet,FC EF,FC Siam conc,FC Siam diff and other deep learning methods,this method is superior to other methods in quantitative and qualitative evaluation,and has great advantages in the accuracy and integrity of the extraction results.For small plaque changes,this method can also effectively identify,and fully verify the efficiency of this method.(4)The domestic high score images of some streets in Binhu District from 2014 to 2015,2015 to 2017 and 2017 to 2019 were detected for building changes.The extraction results of Binhu Century City Street,Yandun street and Yicheng Street are analyzed as a whole,and the changes of buildings in each time period are analyzed by calculating the change compactness.At the same time,the density analysis method is used to analyze the building change density of each street in each time period,and the analysis results are in line with the relevant documents issued by Hefei municipal government.Therefore,this method can provide a theoretical basis for the rational use of urban land resources and future development planning.