Informative Variable Selection in Hyperspectral Images

Abstract

This work analyzes variable selection methods to reduce high dimensionality in hyperspectral images of peanuts contaminated with aflatoxins. A total of 194 spectral bands in the 500–700 nm range were evaluated using statistical techniques, information theoretic measures, group penalized methods, and Relief based algorithms. The results show that bands near 640–700 nm are the most discriminative, while the full spectrum exhibits substantial redundancy. Cross validation with a CART classifier confirms that feature selection improves model stability and reduces overfitting, enabling the identification of informative band subsets suitable for quality control applications in the food industry.