Efficient Wavelength Selection for Limited Near-Infrared Spectral Data via Genetic Algorithm and Hybrid Regression

Abstract

Spectral data often contains a large number of variables that are highly correlated. Although Partial Least Squares (PLS) regression is specifically designed to handle issues arising from limited sample sizes, its effectiveness may still diminish in extremely small datasets, making it challenging to construct a calibration model with high predictive performance. This study introduces a new framework, the Genetic Algorithm and Hybrid Regression Model (GAHRM), designed specifically for variable selection and regression in high-dimensional, low-sample-size spectral datasets. GAHRM integrates Hybrid Regression, which constructs regression models using a covariance structure that is first stabilized through Thomaz Stabilization and then regularized, with Genetic Algorithm (GA), an efficient optimization technique for selecting the best subset of variables among a vast model space. Unlike traditional approaches that rely on exhaustive search for model selection criteria, GAHRM leverages GA to navigate the exponentially large search space, enabling computationally feasible and robust model construction. The effectiveness of GAHRM was validated on the benchmark Gasoline dataset, where it demonstrated superior performance compared to PLS in terms of prediction accuracy and model selection efficiency. These results highlight GAHRM as a powerful alternative for wavelength selection and calibration modeling in challenging data scenarios.