The Research Of Multi-source Remote Sensing Images Change Detection Based On Stacked Denoising Autoencoders

The change detection technology of remote sensing image is the process of quantitative analysis and determinate the change characteristics of objects from remote sensing data of different periods.The main objects of research are natural and artificial objects.With the rapid development of remote sensing technology and constantly emerging of new sensors,the ability which people gain the remote sensing image data has improved continuously.The space remote sensing technology has become from single remote sensing data to multi-source,multi-channel and multi-temporal remote sensing data.Different sensors provide the remote sensing image data of rich and diverse earth's surface.There are consistency and complementarity between these data,how to make full use of these data to dig information which people need is a difficult problem of remote sensing science.In this thesis,we aimed at multi-source remote sensing image change detection,the major works of two aspects are done.The experimental data set is based on synthetic aperture radar(SAR)image,optical image and TM image:1)A new method called stacked denoising autoencoders-based joint classification for multi-source remote sensing image change detection is proposed.Firstly,this approach generates initial change map by two images,and the deep neural networks is used for feature learning for one image,then clusters the image according to the features of learning.Secondly,we select reliable sample pixels from the segmentation of the first image as labels training stack denoising autoencoders for classification ability.Finally,the other image is used as the input of the trained network to automatic classify which achieved the purpose of joint classification.Thus we achieve the purpose of classification,the inconsistent multi-source data types are converted to data types that can be compared directly,then it produces the change map.2)A new method called stacked denoising autoencoders-based feature fusion for multi-source remote sensing image change detection is proposed.Firstly,the log ratio operator is used for processing image to generate the initial change map for the selection of training samples.Secondly,discrete cosine transform,sobel operator and gray level co-occurrence matrix are used to extract the features of two images,then calculatesfeatures difference of two images,and selects reliable training samples to train the stacked denoising autoencoders for multi-feature fusion and classification which are divided into changed and unchanged.Finally,feature difference of all samples are used as the input of the trained network to automatic classify to achieve the final change detection map.The traditional method of remote sensing image change detection generally produces a change map earlier,and then extracts change information from the change map.However,considering the inconsistency of multi-source remote sensing image data,change map contains lots of false change information.It will seriously interfere the analysis and detection.In order to solve this problem,this paper proposes a multi-source remote sensing change detection method which does not extract the change information from the difference map.By simulating the real remote sensing image,the experimental results compare with the traditional change detection method shows that the number of false alarms and missing pixels are reduced and correct rate is improved effectively.