Scene Change Detection By Using Topic Model And Deep Learning For High Resolution Remote Sensing Images

Scene change detection is the process of identifying the differences between the multitemporal image scenes.It could monitor the landscape surface periodically and provide abundant data for the research of society development.There are two kinds of scene change detection methods in remote sensing images: one is the unsupervised scene change detection method,the other is the supervised scene change detection method.Unsupervised scene change detection only judge the scene whether has changed without recognizing the scene.The prior knowledge is not needed in this method,which could be helpful for scene change detection of remote sensing images more directly and automatically.For the supervised scene change detection method,scene recognition is needed to obtain richer semantic information.For unsupervised scene change detection of remote sensing image,a novel unsupervised scene change detection method that based on the topic model is proposed in this paper.The topic model is used for the scene representation firstly.Then,the correlation between the two multi-temporal high-resolution remote sensing image scenes is analyzed in the low-dimensional feature subspace.Finally,the multivariate alteration detection method is used to improve the effect of unsupervised scene change detection.Two experiments show that this method can improve the correlation between unchanged scene images effectively,and improve the result of scene change detection of remote sensing images.For supervised scene change detection of remote sensing image,the deep convolution canonical correlation network is proposed by combining the convolution neural network with deep canonical correlation analysis for supervised scene change detection.Convolution network is becoming the current mainstream of scene representation due to its powerful visual feature expression ability,deep canonical correlation analysis could improve the correlation of data in nonlinear high dimensional space.Therefore,the convolution neural network is used for the scene representation to improve the result of the scene recognition in this network.Then the temporal correlation between two multi-temporal remote sensing image scenes is analyzed in the nonlinear high dimensional feature subspace.Finally,the deep learning network is used for scene recognition and improve the result of scene change detection.Experiment shows that the supervised scene change detection framework could enhance the nonlinear correlation between image scenes in high dimensional feature space,and get more detailed and robust semantic information of scene change detection.