Nearest Feature Line and Point Embedding for Hyperspectral Image Classification

Metric learning methods have been widely used in hyperspectral image (HSI) classification. They can project higher dimensional feature vectors to lower dimensional vectors and get more accurate classification results. Recently, nearest feature line (NFL) embedding (NFLE) algorithm has been proposed in HSI classification. This method tries to embed the distance between a point and its NFL. However, the decreasing of the point-to-line (P2L) distance does not mean that the point-to-point (P2P) distance decreases. In some cases, the P2P distance may even increase, which results in poor classification performance. In this letter, a modified algorithm of NFL and point embedding (NFLPE) is proposed for HSI analysis. Unlike NFLE, which just constrains the P2L distance, NFLPE also imposes an additional constraint on the P2P distance. This additional constraint avoids the possibility that when the P2L distance decreases, the P2P distance increases. Classification experiments with HSI demonstrate its superiority to other related techniques.