Change Detection Of Remote Sensing Image Based On Low-rank Representation And Siamese Network

Remote sensing images are mainly divided into aerial images and satellite images,which contain a large amount of spectral information,and by processing it we can obtain abundant terrains information.As an important research orientation in the field of remote sensing image processing,change detection obtains dynamic changes of surface information by comparing multi-temporal remote sensing images.At present,a large number of change detection methods have been proposed,but their practicability is limited because of the high cost and complexity of manual labeling of supervised change detection methods and the low detection accuracy of unsupervised change detection methods.This paper proposes two remote sensing image change detection methods.Main research contents and innovations include:A new unsupervised change detection method based on low-rank prior regularized discriminative dictionary learning for multispectral images has been proposed.Because of that the changed and unchanged pixels are from different subspaces due to different appearance and statistical property,low-rank representation(LRR)is employed to find informative pixels from the superpixels of difference image(DI).Besides,taking the sparsity of changed pixels in the observed scene into consideration,the selection rule is designed for to distinguish these pixels.Then,discriminative dictionary learning(DDL)is derived for training of changed and unchanged dictionaries from these pixels.Finally,the change map is estimated by comparing the reconstruction error of each pixel in DI on changed and unchanged dictionaries.By LRR,more representative pixels are found for subsequent DDL,which can efficiently improve the performance of the proposed method.Experiments on multi-temporal images from Landsat satellite demonstrate the effectiveness of the proposed method.A new change detection network that based on deep convolutional siamese network has been proposed,which contains supervised pre-training network and supervised finetune network.The former is constituted of two networks with shared weight which called siamese network,by training which it will enlarge the distance of changed pixels at the same location in different multi-temporal images and shrink the distance of unchanged pixels,so the pixels will be easy to classify.Besides,the layers in siamese sub-network are specially designed depending on the characteristics of the input data.The latter is composed of the fully connection layer and softmax layer,which aiming to classify the pixels to get the change map.Experiments on multi-temporal images from Landsat satellite and OSCD dataset demonstrate the effectiveness of the proposed method.