Correlation Alignment Based On Sparse Matrix Transform For Unsupervised Domain Adaptation In Hyperspectral Image Classificatin

Hyperspectral image classification is a research hot spot in the field of hyperspectral image processing.It aims to achieve the classification of unlabeled objects through the model learned from the labeled samples.Due to the difficulty of collection and the high cost of manual labeling,the number of labeled samples is generally very limited.In the case of limited labeled samples,the performance of hyperspectral image classification algorithm model is usually poor.It is a challenging task to realize the classification of ground objects in the case of limited or even no labeled samples.In this paper,domain adaptation(DA)technology is discussed.It is of great practical significance to classify unlabeled hyperspectral data by using known labeled regions or labeled images,so as to reduce the dependence on labeled samples.In this paper,we first introduce the correlation alignment(CORAL)domain adaptation algorithm and discuss its covariance estimation problems in hyperspectral image domain adaptation and classification.Aiming at the instability of covariance estimation in the case of small samples,we propose a sparse matrix transformation(SMT)based correlation alignment algorithm(CORAL-SMT).The correlation alignment algorithm of SMT estimates the covariance matrix of source domain and target domain and obtains accurate and stable covariance estimates.Specifically,the covariance matrix of source domain and target domain is constrained to have a feature decomposition in CORAL-SMT,which can be expressed as a product of a series of Givens rotations.Under the framework of maximum likelihood estimation,the covariance matrix can be effectively estimated by greedy minimization strategy and the estimated covariance matrix can be guaranteed to be positive definite.In the last part of the experiment,this paper tests the proposed algorithm on the data of Gao Fen-5 hyperspectral image and The City of Pavia hyperspectral image,and uses the overall accuracy(OA)and κ coefficient as the evaluation criteria to measure the performance of the algorithm.The experimental results show that the algorithm proposed in this paper has good performance in the test image.