Feature extraction of hyperspectral image based on locally linear embedding

Feature extraction can eliminate the redundant information hidden in the hyperspectral image. It is a necessary preprocessing step of the hyperspectral image analysis system, the classification for instance, to improve the precision and efficiency. Traditional feature extraction algorithms are based on linear transformation, which could not accurately describe the relationship between the original and reduced feature spaces. Therefore, locally linear embedding (LLE), the representative algorithm of nonlinear feature extraction, was adopted to reveal the intrinsic information of the hyperspectral image. For classification, the class labels of the training samples were utilized to adjust the distance matrix and the feature vectors of the test samples were calculated in the way that LLE mapped the unknown samples, realizing the supervised locally linear embedding (SLLE). In the experiment, the nonlinear feature extraction method was combined with three different classifiers and evaluated using the data collected by airborne visible/infrared imaging spectrometer. The experiments show that SLLE is superior to the linear feature extraction methods and can solve the small training set problem of classifying hyperspectral image.