| Remote sensing image change detection aims to detect changes between two images acquired by the sensor of the same geographical area at different time.Crosssensor remote sensing images are obtained by different sensors,which have different representations for the same scene.With the demand of sensor technology and application scenarios,cross-sensor remote sensing image change detection technology has received more attention.However,since the cross-sensor images have different representations for the same scene in a low-dimensional space,it is not possible to directly compare.The deep neural network can represent the complex nonlinear mapping relationship well,map the image to the feature space,and extract the feature information in the image.Therefore,on the basis of a full research on the traditional change detection algorithms,this paper proposes three cross-sensor remote sensing image change detection algorithms based on deep learning.The main work of this paper is as follows:(1)A novel cross-sensor image change detection algorithm based on deep canonical correlation analysis(DCCA)is proposed.This algorithm passes the crosssensor images into the DCCA network pixel by pixel for feature mapping,in order to maximize the correlation of the unchanged pixel pairs in a common latent space,and realize the relative spectral alignment of the cross-sensor remote sensing image.In the latent space,any traditional change detection method,such as change vector analysis(CVA),can be used to obtain the change detection results.(2)A novel cross-sensor image change detection algorithm based on deep canonically correlated autoencoders(DCCAE)is proposed.This algorithm passes the cross-sensor images into the DCCAE network pixel by pixel for feature mapping,in order to maximize the correlation between the unchanged pixel pairs in the common latent space and make the mapped spectral features closer to the original ground object spectral features.Finally,the traditional change detection method can be used to obtain the changes between the mapped features.(3)A novel cross-sensor image change detection algorithm using a hybrid convolutional neural network(HCNN)architecture is proposed.The algorithm uses the Pseudo-Siamese network and the Early-Fusion network to extract the spatial and spectral features of the cross-sensor patches,respectively.Then concatenate the two features to obtain the strong representation spatial-spectral feature of the cross-sensor patches.The feature is put into a sigmoid layer for binary classification to determine whether the feature has changed.In addition,the contrastive loss function is added into the Pseudo-Siamese network to make the unchanged patch pairs closer to each other and the changed patch pairs farther away from each other in the feature space,in order to improve the distinguishing ability of network. |
