A modified correlation alignment algorithm for the domain adaptation of GF-5 hyperspectral image

Abstract

GF-5 is the first full-spectrum hyperspectral satellite used to achieve comprehensive observations of the atmosphere and land. The hyperspectral sensors on the GF-5 have high spectral resolution and wide coverage. However, labeling all the materials on these wide ranges is extremely difficult in practical applications. Hyperspectral classification is extremely difficult when the number of labeled samples is limited or no labeled sample is available. In this study, we aim to present an effective unsupervised domain adaptation technique that uses labeled pixels in the existing old domain(source domain)to classify the scenes with limited or no labeled pixels (target domain). Correlation alignment (CORAL) is a simple and effective domain adaptation method. However, the covariance computation in CORAL may be inaccurate in the case of limited training samples.We propose a new CORAL algorithm on the basis of a sparse matrix transform technique (CORAL-SMT) to solve this problem. The proposed method first uses the sparse matrix transform technique to estimate the covariance matrices of the source and target domains and then performs the CORAL between the estimated covariance matrices. The SMT method can obtain an accurate covariance estimator, which is constantly positive and definite. In the experiment, we compare the proposed CORAL-SMT with some classical domain adaptation methods, such as subspace alignment, principal component analysis, CORAL, transfer component analysis, geodesic flow kernel, and information the oretical learning. After domain adaptation, we use the nearest neighbor and support vector machine as classifiers to classify the unlabeled data in the target domain. Two GF-5 hyperspectral datasets, namely, Huanghekou and Yancheng datasets, are used to evaluate the performance of different methods. Experimental results demonstrate the effectiveness of the proposed method compared with subspace-based alignment methods and CORAL. The GF-5 data have good spectral discriminative ability. In the case of bias sampling, the performance of classifying the target samples on the basis of the source model is acceptable. The distribution difference between source and target domains is decreased, and the classification performance is intensively improved using the domain adaptation technique. The SMT technique can improve the covariance estimation,thereby benefiting the following domain adaptation.