Hyperspectral Remote Sensing Image Classification Based On Semi-supervised Support Vector Machine

There are so many traditional technologies applied in remote sensing imageprocessing outing of action, confronted with the characteirstics of the hyperspectralimage such as high demensions, strong correlation among the bands. For thehyperspectral classification, the other challege is that there are few labeled samples tobe used. The semi-supervised learning takes andvatage of ulabeled samplesinformation to help the labeled samples to train a classifier with a better performance,which focuses on sloving the ill-conditioned problem caused by few labeled samplesused in model training. This thesis has appled the semi-supervised methodintergrating the support vector machines to hyperspectral iamge classification, andbased on the systematic research of semi-supervised theories, studied the problems inhyperspectral image classification based on semi-supervised support vector machines,and proposed the corresponding solutions. The main contirbutions and conclusions ofour works in this dissertation are summaried as fellows:(1)In view of the the disadvantage of selection strategies used in traditionalself-training that the unlabed samples are selected based on the distance between aunlabeled samples and one class, which always results in unbalance of trainingsamples, we proposed a strategy that conducted by the distance between two samples,then established an semi-supervised classification algorithm for hyperspectral imagebased this strategy and support vector machines. Two real hyperspectral remotesening images have been used to evaluate the algorithm and the experimental resultsindicate that the performance of the algorithm is stable and it effectively improved theperformance of support vector machines classfier trained by few labeled samples.(2)The spatial information has been applied for the process of unlabeledsamples selecetion, and it successfully solved the error accumulation because of theunfitness of traditional similarity meansure approaches used in the hypersepctraliamge processing. Meanwhile, the spatial information makes the training sampleshave more information of the space distirbution, and it increases the information oftraining samples as well. Experimental results indicate that the proposed algorithmhas effectively improved the overall accuracy of hyperspectal image classification.(3)A new co-training algorithm based on single view and mutli-classifer hasbeen proposed. The two calssifiers, support vector machines and k-NearestNeighbour, have been employed to construct the algorithm. The expeirmental results show a better perfoermance in calssification, and indicate that the co-training methodhas advantages in inconsistent infromation mining and solving the problem ofinsufficient information in the semi-supervised classification process.