Research On Remote Sensing Image Change Detection Method Based On Spatial-spectral Feature Combined Sparse Representation

As the basic work of geographical conditions monitoring,Remote sensing change detection provide scientific decisions for the urban development.The change detection of remote sensing images is the process of determining the change information according to several observations in different time,then the extracted information can be evaluated.Due to the result obtained by single feature(i.e.spectral)change detection method is not complete,while the different features can represent the image from various views,furthermore the fusion of multi-feature is able to enhance the ability to identify the objects,the accuracy of image detection can be improved.Based on the above considerations,this paper will carry out further research on the change detection method around the fusion of spatial-spectral features.The contents and main achievements are as follows:(1)For the traditional single weight spatial-spectral feature combination,each pixel in the spectral and texture change magnitude images was given a fixed weight,and it cannot adapt to the changes in the image,a double weight fusion strategy of spectral and texture features was proposed.The weight was defined by the difference in expression of between homogeneous and non-homogeneous regions in a remote sensing image.The steps were as follows:Firstly,spectral and texture change magnitude images were produced respectively,and meanwhile the double-weighted model was designed;Secondly,the final spatial-spectral fusion image was produced by the model;Finally,the result was determined by the Otsu binary threshold algorithm.The experiment of Quickbird image showed that the method has achieved better performance(Overall Errors=6.49%,Overall Accuracy=93.51%)than Spectral only Change Vector Analysis method(SCVA)(Overall Errors=12.10%,Overall Accuracy=87.90%)and Single weight Change Vector Analysis method(SWCVA)(Overall Errors=9.44%,Overall Accuracy=90.56%).(2)A new change detection method of spatial-spectral feature double dictionary cross sparse representation was designed.This method was based on the traditional feature fusion weighting method,firstly,the dictionary was obtained by training samples,and sparse representation was used to construct spectral and texture change magnitude images respectively,in which sparse representation reduced the uncertainty of results by adopting cross representation.Secondly,the spatial-spectral feature difference image was combined by weight.Finally,the K-means clustering method was used to forming a binary change detection map.The experiment of Quickbird image showed that the method improved the accuracy greatly(Overall Errors=4.32%,Overall Accuracy=95.67%).(3)A change detection method of multi-task sparse representation with spatial-spectral information was proposed.Firstly,the sparse representation of image under a single feature was regarded as a task,and the representation of each feature channel was obtained through the sparse representation model.The sparse representation coefficients of different features had similar structure,and the spatial neighborhood information of pixels were combined at the same time;Secondly,the joint dictionary set was obtained by training samples,and the joint sparse representation model with multiple features was constructed.The experiment of Qucikbird image demonstrated that the method can fully exploit the relationship between different features,making the results more accurate.(Overall Errors=3.20%,Overall Accuracy=96.79%)...