Research On Change Detection Based On Autoencoder And Level Set Method

Change detection is a process that identifies surface changes in different periods in the same area,which is widely utilized in land expansion,agricultural monitoring,disaster assessment and other fields.The progress of remote sensing and space technology has brought more abundant satellite images,which provides opportunities and challenges for accurate and real-time extraction of change information.For the low and medium resolution satellite imagery that is extensively exploited in the practical problems of change detection,it is necessary to improve the accuracy and degree of automation of change detection.With regard to the high-resolution satellite image whose spatial features are rich,the phenomenon of “the same objects with different spectrum” and “the different objects with same spectrum” is prone to appear in the images,so the key tasks are to reduce the redundant features of the images and improve the change detection accuracy.This paper focuses on the research of change detection based on autoencoder and level set evolution method and applies methods to land use,reclamation and coastline change detection,which increases the accuracy of change detection and promotes the degree of automation of change detection.In the aspect of land use change detection research,for manual screening of ground-truth samples and the problem of imbalance between the changed and the unchanged samples,a change detection method based on collaborative clustering and weight-attention sparse autoencoder are proposed in this paper.Cooperative clustering and synthetic minority oversampling technique methods are exploited for generation of the balanced samples unsupervised.Then the weight-attention mechanism is introduced into the autoencoder to improve the accuracy of classification.Finally,the “salt and pepper” noise of change detection results is reduced by the spatial constraints based on the superpixel segmentation.To reduce the impact of suspended sediment on the extraction of reclamation change information,a joint change detection method based on weight-clustering sparse autoencoder is proposed in this paper.The mean and variance of superpixel blocks are stacked as training samples for acquiring the spatial neighborhood information of ground objects.The redundant features are reduced by merging similarity weights of autoencoder layer by layer.Finally,the change detection map is generated by the joint probability judgement combined with the advantages of difference map classification and object-oriented change detection.Experimental results show that the proposed method significantly decreases the false detection rate and omission detection rate and improves the accuracy of object-oriented change detection.In the aspect of coastline change detection,this paper focuses on the automatic extraction of coastline change information and proposes a coastline change detection framework based on the improved distance regularization level set evolution method.Firstly,the clustering results of normalized difference water index extracted from remote sensing image are processed by morphology methods for generation of the initial level set function,and then the local entropy of image is mapped nonlinearly to initialize the improved distance regularization level set method.By introducing the exponential edge stop function based on posterior probability,evolution curves can rapidly evolve to the coastline in the proposed method.The final coastline change detection results are obtained by comparing and analyzing the coastline results of bi-phase images.Experimental results show that the proposed method can extract satisfactory coastline results with same or fewest iterations,and the coastline change detection map is highly consistent with the ground-truth map.