Remote Sensing Image Change Detection Based On Self-Paced Learning And Stacked Denoising Autoencoders

Remote sensing image is important data observing changed situations on the earth,we can effectively identify specific change areas by image change detection technology.Change detection,a process distinguishing the changed areas and the unchanged areas in two remote sense images of the same region but taken at different points in time,has been adopted in many application fields.Recently,the research focus has been developed from the unsupervised change detection method to the change detection method based on supervised classifier.Many proposed change detection methods based on supervised classifier obtain good performance,however facing to the increasingly complex image data,we need continue researching and innovating for further improving performance of change detection method.In this paper,SAR image change detection and multi-spectral image change detection are researched.Specifically,we propose a change detection method based on self-paced learning and stacked denoising autoencoders,then work on improving the proposed method.Our research achievements are described as follows:1)Propose a SAR image change detection method based on self-paced learning and stacked denoising autoencoders(named self-paced stacked denoising autoencoders model).This proposed method aims at solving the problem existed in change detection methods based on supervised classifier,i.e.,how to help supervised classifier learning reliable data.Specifically,we adopt stacked denoising autoencoders as the supervised classifier and adopt selfpaced learning to help stacked denoising autoencoders optimizing its objective function.In the iterative optimization process,self-paced learning assigns samples various weights to help stacked denoising autoencoders learning progressively reliable sample and improving its classification performance.Additionally,we adopt differential evolution algorithm to optimize a pace parameter sequence used in self-paced learning for further improving the change detection performance of proposed method.2)Propose a SAR image change detection method based on ensemble learning of self-paced stacked denoising autoencoders model.On the basis of previous work,this method introduces ensemble learning mechanism.Specifically,we adopt several sampling methods on the original dataset to obtain various training datasets at first;then implement proposed selfpaced stacked denoising autoencoders model on every training dataset,which can form a learning grid for stacked denoising autoencoders(vertical direction: an ensemble learning process based on different training datasets,horizontal direction: an iterative optimization process based on self-paced learning);we can obtain the final change detection result by combining the change detection result of every base learner(i.e.,self-paced stacked denoising autoencoders model)for improving the change detection performance of previous work.3)Propose a multi-spectral image change detection method based on self-paced learning and classification information.While implementing proposed self-paced stacked denoising autoencoders model on multi-spectral image,this proposed method considers the natural classification information of multi-spectral image data.Specifically,in the optimization process of self-paced stacked denoising autoencoders,we divide original dataset into several groups according to sample’s natural class and adopt self-paced learning to update sample’s weight in every group respectively.This way can help stacked denoising autoencoders learning reliable and diverse data to train its classification performance.Additionally,we adopt a sample feature extraction way of multiband data permutation and combination to help proposed method learning effective feature representation.