Modified Nonparametric Weighted Feature Extraction Algorithm

Nonparametric weighted feature extraction (NWFE) has been proven to be a powerful feature extraction tool for hyperspectral data classification with a weight function based on Euclidean distance (ED). In this paper, we propose a modified algorithm referred to as nonparametric weighted spectral pan-similarity measure feature extraction (NWSPMFE). In NWSPMFE, ED is replaced by the spectral pan-similarity measure, and the weight function is redefined in scatter matrices for NWFE. The performance of NWSPMFE is evaluated by comparing it with principal component analysis (PCA) and NWFE in terms of overall accuracy and Kappa analysis based on two experiment datasets. The overall classification accuracies of PCA, NWFE, and NWSPMFE for D.C. Mall and Indian Pine datasets are 0.942, 0.949, 0.961 and 0.496, 0.665, 0.697, respectively. However, NWSPMFE’s runtime is slightly longer than that of NWFE.