Building Change Detection In Remote Sensing Image Based On Convolutional Neural Networks

Building change detection is the process of identifying variations of buildings by observing two images of one place taken at two different periods.The variation of change information includes the increase or decrease.This research has great social significance and high economic value in the domain of social studies such as town and country construction,land resource utilization,disaster control and assessment,etc.The technology of deep learning has progressed at a high rate of speed,and many excellent deep convolutional neural network models and their variants have been proposed.However,most existing building change detection methods treat the bi-temporal images as the input of the network,ignoring joint spatial information.And the change detection process is divided into two steps: building information and change identification,which adds the error accumulation.In order to solve these problems,this thesis accomplishes the two tasks simultaneously,instead of extracting building information and discriminating change areas separately.This thesis researches the building change detection in aerial images from the following steps:(1)To extract joint spatial information from the bi-temporal input images,this paper combines the two 3-channel bi-temporal images into a single 6-channel image as the input of these networks.The feature needed to be learned and extracted is double.Furthermore,two multi-feature units of two-branch and three-branch are proposed.Based on the different ways of using multi-feature units,MF-Net and MFH-Net are proposed.In MF-Net,two multi-feature units are symmetrically used in the encoding and decoding part to extract different scales of pixel information.Compared with other detection methods,the accuracy of this model has been improved.(2)However,the computational cost of MF-Net is very high.By decreasing the use of multi-feature units and using the feature maps from center and decoding part to compose the head part,MFH-Net is further proposed.(3)In the MF-Net and MFH-Net,the feature maps obtained from the multi-scale feature unit are directly used without information filtering and screening.The thesis does some experiments about the import of attention modules after three-branch multifeature unit in the encoding part and proposes MFHA-Net.Furthermore,the thesis combines the multi-feature unit with channel and space attention to get a new model named Channel Multi-Feature Space Attention Module(CMFS).Thus,a new neural network MFHA-Net+ is proposed.The MFHA-Net+ could not only extract the joint spatial information of the bi-temporal images from different layers and pay attention to the change information,but also make use of the feature maps from the center and decoding part for further feature fusion.A comparative experiment is carried on the remote sensing building image WHU data set.The evaluation metrics we use in this study are precision,recall and F1 value.The results show the superiority of our proposed methods,compared with other network models such as FC-EF,FC-Siam-conc,FC-Siam-diff,etc.And the F1 value has achieved 92%.