Hyperspectral Remote Sensing Image Classification Based On Ensemble Learning

With the rapid development of hyperspectral imaging spectroradiometer of highspatial and spectral resolution, the existing data processing methods face a newchallenge and we require new data processin algorithms. As one of the most effectivemethod of machine learning, ensemble learning, which utilizes many learningalgorithms to solve a predition problem, can significantly improve the accuracy andstability of a learning system. This dissertation introduces ensemble learning intohyperspectral remote sensing image classification. The main contributions of ourworks in this dissertation are summarized as follows:1) Rotation Forest has been applied for hyperspectral remote sensing imageclassification. Based on the framework of original Rotation Forest, we proposeRotation Forest with different feature extraction algorithms, such as independentcomponent analysis, maximum noise fraction and local fisher discriminant analysis.Experimental results indicate that the performances of Rotation Forests are better thanother tradational ensemble learning algorithms (Bagging et al), especially withprincipal component analysis and independent component analysis feature extractionmethods.2) Supervised/semi-supervised principal component analysis has been used toextract features of hyperspectral remote sensing image classification. Theperformance of supervised/semi-supervised principal component analysis is comparedwith other traditional feature extraction methods and evaluated based on severalcriterias: different datasets, different number of labeled/unlabeled samples and thecomputation time. Experimental results revealed that supervised/semi-supervisedprincipal component analysis can extract more reliable features for hyperspectralimage classification than other feature extraction algorithms. Futhurmore,parallel/concatenation ensemble methods based on three powful supervised/semi-supervised feature extraction, including supervised/semi-supervised principalcomponent analysis, non-parametric weighted feature extraction, are proposed forhyperspectral image classification. Experimental results show that the ensemblemethods can improve the accuracy of hyperspectral remote sensing imageclassification.3) Spectral-spatial ensemble method based on semi-supervised feature extractionand image segmentation or markov random field are developed for hyperspectral image classification. We choose clustering, watershed transformation and mean-shiftas image segmentation techniques. The clustering algorithms include K-means,ISODATA, Fuzzy K-means, Kernel K-means and EM. Markov random field adoptsthe simulated annealing algorithm to make the adjacent pixels aggregation byminimizing the local energy function. The above spectral-spatial methods significantltimprove the classification accuracies and reduce the noise of classification results.4) Based on the theoretical research, hyperspectral image ensemble learningclassification system (HIELCS), which includes image preprocessing, classification,clustering, segmentation, train/feature stage ensemble learning, spectral-spatial,diversity measures et al, is implememted. Urban impervious surface area extractionusing HIELCS sytem shows the advantages of hyperspectral remote sensinginformation processing in practical application.