| In the analysis of remote sensing image,image features are the basis.And the important step before remote sensing change detection is that how to scientifically and effectively select features that are beneficial to the construction land cover change detection,avoiding the decrease of operation efficiency and detection accuracy in subsequent change detection process as a result of feature redundancy.Considering the difficulty of obtaining label information in remote sensing images,unsupervised feature selection methods present more research significance and application value.Based on a typical embedded unsupervised feature selection method,MultiCluster Feature Selection(MCFS)algorithm framework,this paper proposed an improved Multi-Cluster unsupervised feature selection method(IMCUFS)for feature selection requirements from remote sensing images,which addressed the problems of high computational complexity,multi-parameter tuning and feature vector selection in the MCFS algorithm.For algorithm verification,IMCUFS was applied to remote sensing construction land change detection.Two remote sensing data sets were selected for method comparison and validation.The first set of data was Triple Sat-2 multi-spectral images in Nanjing on November 27,2016 and July 18,2017,and the second set of data was World View-2 images in Nanjing on September 16,2013 and July 29,2015.Used BJ-2images as the first group of study area images to verify the feasibility and effectiveness of our method.Using 6 typical feature selection methods were selected as comparisons,in which,three supervised feature selection methods are Relief F,Fisher score and Minimum Redundancy Maximum Relevance(MRMR),and three unsupervised feature selection methods are Laplacian Score(LS),MCFS and Unsupervised Discriminative Feature Selection(UDFS).Based on K-means unsupervised and support vector machine(SVM)supervised classification methods,accuracy was compared while using different feature sets,including original spectral change vectors,full features extracted,the features obtained by our method and other six methods.To further verify the applicability of our method,WV-2 images were used as the second group of study area images.Based on the K-means and SVM methods,accuracy was compared while using different feature sets,including original spectral change vectors,full features extracted,the features obtained by our method and other six methods.The main conclusions of this paper are as follows:(1)In the experimental and comparative verification based on BJ-2 images,the feature extracted in this paper achieves the highest detection accuracy.Compared to the change detection accuracy of the features selected by our method and six typical feature selection methods,the results showed that the average change detection accuracy of the proposed method is 73.13%,which is 1.65% slightly than the full feature,and the highest accuracy of unsupervised change detection based on k-means.In the supervised SVM change detection,the average detection accuracy achieved by our method is 81.68%,which is 0.59% slightly lower than Relief F algorithm,and slightly higher than 0.19% of full features.And the change detection result of our method is the best among all reference methods.(2)In the application verification based on WV-2 images,the feature extracted in this paper achieves the highest detection accuracy either.In the unsupervised change detection based on K-means,the average accuracy of the change detection of our method was 64.62%,which was 19.68% beyond all features,and far higher than the other feature selection methods.In the supervised change detection of SVM,the average accuracy of our method was 71.51%,which is 0.09% slightly higher than all features.Compared with the reference methods,the performance of our method was the most stable.In conclusion,our method can effectively select a small number of image features,and obtain higher change detection accuracy.Under the same conditions,Average accuracy and stability of change detection results are better than that of other six feature selection methods.This shows that proposed method can be applied to the feature selection of construction land change detection in high-resolution images,so as to provide unsupervised feature selection for remote sensing change detection. |
PDF Full Text Request