LPLDE (locally preserving linear discriminant embedding) method of feature extraction for hyperspectral image classification

In view of the difficulty to get enough training samples in hyperspectral image, this paper presents a novel supervised linear manifold learning feature extraction method based on the manifold learning, Fisher criterion and Maximum Margin Criterion, named LPLDE by us, for hyperspectral image classification with nearest neighbor (NN) classifier. The intraclass compactness and interclass separability of hyperspectral data are respectively characterized by within-class neighboring graph and between-class neighboring graphs via embedding. The LPLDE method which efficiently avoids the within-class scatter matrix singularity caused by small-sample-size problem has better discriminative performance and is more suitable for classification. Experimental results on hyperspectral datasets and their analysis demonstrate preliminarily the efficiency of our LPLDE method as compared with other existing methods.