Research On Hyperspectral Image Classification Algorithms Based On Spatial-spectral Joint Model

Hyperspectral image(HSI)classification is a very important research topic in remote sens-ing,which has many realistic applications.How to jointly consider the spectral feature and spatial information of a given testing pixel is a critical issue for improving the performance of HSI classification algorithms.This thesis mainly focuses on studying the HSI classification methods based on the spatial-spectral joint model.The main contributions of this thesis are as follows:First,this thesis presents a novel robust joint nearest subspace(NS)model,which extends the original NS model to solve the HSI task.This model introduces the additive sparse noise into the the joint NS model for considering the outliers within the neighboring pixels of the testing sample,and then defines a robust NS distance to measure the difference between the neighboring pixels and the training set of each class.Besides,this thesis develops a contextual prototype learning algorithm to learn effective prototypes regarding the training samples.The experimental results demonstrate that the proposed robust NS distance and the contextual prototype learning method could improve the classification accuracies step by step.Second,this thesis develops a novel HSI classification framework based on two stage joint representation(TSJR),which comprehensively consider how to effective obtain the represen-tation coefficients and how to further exploit the representation coefficients for classification.For one thing,this thesis proposes the joint collaborative representation model and joint struc-tured sparse representation model,and develops the related optimization methods.For another,this thesis considers the obtained coding coefficients as meta-features and uses them to train a classifier.In the testing process,the TSJR framework combines the outputs of both joint rep-resentation model and classifier to conduct the final classification.The experimental results show that the proposed framework performs better than other competing ones and has better generalization ability.Third,this thesis designs a novel HSI classification framework based on spatial-spectral K nearest neighbors,which introduces the original KNN method to the HSI field by defining a set to point distance.The major contribution is to propose a novel set to point distance in terms of a least squares manner to depict the distance between the neighboring pixels and each training sample.From the view of subspace,this distance measures how far each training sample from the subspace spanned by the testing pixels.By combining the aforementioned distance and the weighted K nearest neighbors method,the proposed classification framework achieves better performance.Finally,this thesis proposes a novel HSI classification method based on set-to-set distance.This method treats the neighboring pixels of the testing sample and the training samples from each class as two individual convex hulls,and then exploits the convex hull representation model to depict the set-to-set distance.Compared with other HSI classifiers proposed in this thesis,this method is able to simultaneously model the spatial-spectral information of the neighboring pixels of the testing sample and the structure information of training samples from each class.In addition,we develop a novel label optimization method based on energy minimization,which could further optimize the classification probabilities output by the traditional HSI algorithms and significantly improve the classification accuracies of those methods.