Researches On Hyperspectral Images Classification Technique Based On Random Walk Optimization

Hyperspectral imaging can capture more than one hundred spectral channels. Therich spectral information preserved in the hyperspectral images makes hyperspectralimages more suitable for discrimination and detection of the land-cover object typesthan remote sensing images because of their abilities of characterizing the materialsof objects on the surface of the earth. Therefore, hyperspectral images classificationhas become a hot research topic in remote sensing application. However, the highspectral dimension of hyperspectral images also brought problems to hyperspectralimages classification, such as redundancy between adjacent bands and “Hughes”phenomenon. Recently, studying an efficient and effective hyperspectral imagesclassification technique has been investigated by many researchers. This thesis focus-es on integrating spatial information into support vector machines (SVM) classificati-on of hyperspectral images based on SVM and random walk (RW), aiming at solvingthe problems such as insufficient training samples, long algorithm running time andthe “Hughes” phenomenon caused by numerous spectral bands. In this thesis, twoclassification techniques based on RW have been proposed to combining spatialinformation with spectral information. The main work is briefly summarized asfollows:1. Hyperspectral images classification technique using random walk based supe-rpixels (RWS) segmentation: This hyperspectral images classification technique isproposed for integrating spatial neighborhood information into the hyperspectral ima-ges classification to get an improved classification result. Firstly, the classificationmaps are obtained by SVM classifier. Secondly, the random walk based segmentationmethod is conducted on the decomposed hyperspectral images obtained by princepalcomponent analysis (PCA) to get a precise segmentation result. Finally, the resultingclassification map is obtained by performing major voting on the SVM classificationmaps and the RWS segmentation results.2. Hyperspectral images classification technique based on RW optimization: RWoptimization aims at optimizing the initial classification probabilities obtained bySVM with the spatial neighborhood information of hyperspectral images. For thistechnique, the SVM classification probabilities are treated as an aspatial energy term.The spatial connecttednesses between adjacent pixels are modeled as a spatial energy term. Based on the RW algorithm, the classification maps can be obtained by minizingthe combined energy function. The final classification results indicate that classificat-ion technique based on RW optimization can also improve the SVM classificationaccuracy effectively.To demonstrate the effectiveness of the proposed techniques, experiments areperformed on the Indian Pines and Salinas databases. Experimental results indicatethat the two hyperspectral classification techniques based on RW can achieve verygood classification performances in terms of classification accuacies, which meansthat RW algorithm indeedd works in solving the problems of large data, high datadimensions and data redundancy. In particular, the classification technique based RWoptimization can acquire a remarkable classification accuracy in terms of limitedtraining samples.