Relaxed Clustering Assumption Based Hyperspcctral Image Classification Methods

In recent years, hyperspectral remote sensing has been rapidly developed. Hugehyperspectral data has been got, while how to discover the interested information fromthe huge data is to be resolved. As an important means of information mining,hyperspectral data classification has been a hotspot. However, a large number of mixedpixels in hyperspectral data seriously restrict the improvement of classification accuracy.To address these issues, the relaxed clustering assumption is cast on the hyperspectralvectors to formulate a modified spectral similarly regularizer in the spares coding andthe semisupervised support vector machine. Then combing it with spatial laplaceregularizer, a few good hyperspectral data classification methods are proposed.The maincontents and innovations are as follows:(1) A hyperspectral data classification method based on relaxed clusteringassumption and sparse coding is proposed. The relaxed clustering assumption isextended to the sparse coding model in a probability vector form. A new objectivefunction is obtained. Then optimize it, we can receive the itreative formula.Theeffectiveness of the proposed method is evaluated on hyperspectral data. Theexperimental resultes show that the relaxed clustering assumption can greatly mitigatethe influence of mixed pixels on the classification accuracy.(2) A relaxed clustering assumption based semisupervised hyperspectral dataclassification is proposed. In this method, the relaxed clustering assumption is extendedin the semisupervised support vector machine. This method firstly projects all thesamples into the feature space by the kernel function, then classify the projectedsamples. Based on the properties of hyperspectral data, a laplace graph is calculated forclassification. Since the relaxed clustering assumptin reduces the possibility ofmisclassification and the laplace graph can smooth the labels to avoid thesalt-and-pepper noises in the classification map, higher classification accuarcy can beobtained. Some experiments on several real hyperspectral data are taken to prove theeffectiveness of the algorithm. Compared with the available methods, this algorithmexhibits more accurate classification of hyperspectral vectors, even with small-sizelabeled samples.(3) A spatial-spectral relaxed clustering assumption based semisupervisedhyperspectral data classification is proposed. The RCA-SLR-SSC is further relaxed inspatial. First of all, the sketch method is used to extract the pixels at the boundary between the background and the pixels to be classified. Then reduce the correspondingweights in spatial constraint matrix. A new spatial constraint matrix is got. Then thematrix is used in RCS-SLR-SSC. The spatial relaxed clustering assumption can greatlymitigate the misclassification of pixels at the boundary and increase the classificationaccuracy. The experiments on real hyperspectral data certify that this algorithm hasobvious advantages compared with RCA-SLR-SSC on classifying the pixels at theboundary.This work was supported by the National Basic Research Program of China (973Program) under Grant no.2013CB329402, NCET-10-0668, National ScienceFoundation of China under Grant no.61072108,60971112,61173090, The ministry ofeducation doctoral program funds (20120203110005), Weapons and equipmentadvanced research fund project (9140A24070412DZ0101) and Higher school subjectinnovation engineering plan (111plan), No. B0704.