AN IMPROVED SUPERVISED RANKING STRATEGY FOR FUNCTIONAL PRINCIPAL COMPONENTS IN HYPERSPECTRAL IMAGING

Abstract

Unlike conventional imaging systems, which capture data in discrete RGB channels, hyperspectral imaging (HSI) technology provides detailed spectral information across a continuous wavelength range. Analyzing HSI data presents a challenge due to its infinite-dimensional feature space and relatively limited sample size. A natural and effective approach to handling such functional data is functional principal component analysis (FPCA), which serves as a dimension reduction tool and an extension of traditional principal component analysis (PCA). In standard FPCA, the leading functional principal components (FPCs) are typically ranked based solely on the proportion of explained variance. In this paper, we propose an improved supervised ranking strategy for the leading FPCs that prioritizes discriminative power. We utilize two machine learning algorithms–support vector machines (SVM) and random forests (RF)–to inform this ranking. After applying this dimension reduction technique, the extracted, highly discriminative features are input into machine learning algorithms for downstream statistical analysis. We demonstrate the superior effectiveness of our proposed methods using two real-world hyperspectral datasets and two simulation studies.