Study On Semi-supervised Classification For Hyperspectral Remote Sensing Images

In recent years,hyperspectral remote sensing images have received great attention because of its many advantages,such as high resolution,continuous spectral curves and more abundant information of ground objects.The characteristics of hyperspectral data,many bands and the strong correlation between adjacent bands,lead to dimension disaster and Hughes phenomenon.It means that remote sensing data contain a lot of redundant information,and then the classification results of land cover are affected.Therefore,it is an important topic to extract useful knowledge from hyperspectral remote sensing data,while reducing the dimension of hyperspectral data and keeping the useful information in the original band as much as possible.For hyperspectral remote sensing images,labeling samples is not only time-consuming,but also on the experience and knowledge of the experts.Semi-supervised learning can make use of a small number of labeled samples and some unlabeled samples to form a training sample set and a test sample set respectively.The performance of the classifier can be improved by learning training.Generally speaking,the classification accuracy is proportional to the number of labeled samples.Therefore,it is a challenging task how to use limited training samples to obtain ideal classification results.On the basis of summarizing and analyzing the existing results,this dissertation has carried out the following research work.(1)Based on the spectral method,spatial information and genetic algorithm,this dissertation proposes a new classification method for hyperspectral image.A new similarity function is firstly defined by the spectral information and the spatial information of the samples.Then an optimal weighted graph will be obtained by using of genetic algorithm to optimize the original graph which is constructed by k-nearest neighbor method.Based on the optimization graph,the improved spectral method is used to reduce the dimension of hyperspectral data.Finally,the hyperspectral image classification is completed by the local mean pseudo-nearest neighbor method.(2)Based on pre-classification and discontinuous probabilistic relaxation(DPR)strategy,a novel semi-supervised classification algorithm for hyperspectral images is developed in this dissertation.Firstly,DPR and Robert operators are used to denoise the original hyperspectral data.Based on the training set of limited samples,two classifiers,MLRsub and LMPNN,are applied to pre-classify the denoised images to achieve the purpose of expanding the training sample set.With the help of the new training sample set,the MLRsub classifier is adopted again to classify hyperspectral data.In order to obtain the good classification accuracy,the DPR strategy is employed to improve the classification map in post-processing.In order to verify the effectiveness of the proposed algorithm,we have carried out experiments on three typical datasets,Indian pines,Salinas and Botswana,which are widely used to test the completeness of the hyperspectral classification algorithm.The experimental results and the comparative experimental results show that the proposed method can obtain high classification accuracy.