Gaussian Processes For Hyperspectral Images Classification

With higher spectral resolution than the traditional remote sensing approaches,hyperspectral remote sensing can provide abundant spectral information for earth observation,and thus plays an increasingly important role in national economy and modern military fields.However,during the processing of the hyperspectral imagery,several new problem and challenges are resulted from the high dimensional data sources.For example,how to classify high-dimensionality data with limited availability of reference data,how to deal with the inter-class spectral similarity between different classes,and how to collect enough labeled set with high quality to obtain a good classifier.Gaussian process(GP)model,which was recently proposed at the end of 20 th century,is a novel kernel method in machine learning.Besides the common advantages in processing high-dimensionality data with kernel methods,it also has many other merits.GP model is presented with fully Bayesian formula,and can be implemented easily via the perfect mathematical properties of Gaussian distribution.And moreover,another interesting merit of GP model is that the hyperparameters could be chosen automatically.Based on the deeply analysis of the characteristics in hyperspectral imagery and systematically review of the state-of-the-art classification methods,researches on GP model for classification of hyperspectral images are conducted in this dissertation,and multiple novel hyperspectral image classification methods are proposed.First,classification method with GP regression(GPR)model is studied,in which only the spectral information is used to obtain the classification results.To combine the spectral and spatial information for classification of the hyperspectral image,a new spectral-spatial GPR(S2GPR)model is proposed.In the proposed S2 GPR method,the classification results are refined by the labels of neighborhood pixels based on the classification results of the original GPR model.And the cosine similarity measure is introduced in S2 GPR to combine and weight the contribution of the two kinds of information.In this way,the classification results of original GPR model using only spectral information could be improved by the consideration of the spatial constraint of neighborhood labels.Compared with results of the original GPR model,classification accuracy with the proposed S2 GPR method on the AVIRIS Indian Pines and ROSIS Pavia University data sets could be improved by about 6% and 5%,respectively.Beside the S2 GPR method to incorporate spatial information,the model updating issue when new labeled samples are added to the training set is also studied,and a new model updating scheme based on Cholesky matrix decomposition is proposed,which could update the GPR model in an efficient way.Then,GP classification(GPC)model,which could provide probabilistic outputs,is studied for classifying hyperspectral imagery.To improve the classification results with GPC model which could only use the spectral information,a spectral-spatial classification method which combine Markov random fields(MRF)and GPC model is proposed.In the proposed GPC-MRF method,the probabilistic results are provided with GPC,and MRF is used to model the high-level contextual relationships in spatial domain,and the final classification results are obtained with the maximum a posterior(MAP)decision rule under Bayesian framework.Experiments are conducted on two typical hyperspectral images: AVIRIS Indian Pines and ROSIS Pavia University data,and the classification accuracy with the proposed GPC-MRF method with respect to the original GPC model could be improved by about 10% and 6%,respectively.The results show that GPC-MRF could greatly improve the classification performance of the GPC model,even with a small amount of training sample.And then,to decrease the possible classification error with a single classifier,the multiple classifier system(MCS)is introduced to classify hyperspectral images.Considering that redundancy is usually existed between the high dimensional spectral features,a classifier ensemble method with GPC model is proposed with the partition of the entire spectral bands in the hyperspectral image.In the proposed method,multiple GPC models are constructed on different data sources resulted from the spectral partition step,and the final classification results of the test data are obtained by the fusion of the binary probabilistic outputs with each GPC model.Moreover,the GPC-MRF is adopted to further refine the results after the classifier fusion step.The proposed method could to some extent alleviate the imbalance problem between the high dimensional features and limited availability of training samples.From the experimental results,we can conclude that the classification results with a single GPC could be improved with the proposed method,if the step of spectral partition could result in a diverse set of classifiers.Especially on the AVIRIS Indian Pines data,the overall accuracy with the proposed spectral-spatial classification method via the classifier fusion step with respect to the GPC-MRF method could be improved by about 7%.Finally,to deal with the problem of limited availability of training samples in hyperspectral image classification,active learning(AL)with GPC for iteratively collect the training samples from unlabeled set is studied.Based on the posterior probabilities provided by the GPC model,three different sample selection heuristics are proposed to actively select the candidate samples in the unlabeled set,and then labeled by an expert.Begin with a small number of labeled set,the proposed AL heuristics could effectively select the informative and useful samples from the unlabeled candidates,leading to a fast increase of the discriminative ability of the GPC model.In addition,to speed up the model updating procedure during the AL process,an incremental-like model updating strategy is proposed.With the proposed strategy,the unnecessary retraining of GPC model could be avoided during two sequential AL steps,and thus saving the computational cost.Experimental results on two real hyperspectral images demonstrated the effectiveness of the proposed AL heuristics and the model updating strategy.