Advances in Machine Learning Framework for Near-Infrared Spectroscopy: A Taxonomic Review on Food Quality Assessment

Abstract

This review taxonomically analyzes and evaluates recent advances in machine learning (ML) frameworks applied to near-infrared spectroscopy (NIRS) for food quality assessment. Through a comprehensive literature search across IEEE Explore, ScienceDirect, and Springer (2021-2024), we examine key framework components: data acquisition, public datasets, preprocessing, wavelength selection, and advanced ML architectures. Our analysis reveals the current state: miniaturized devices and multi-device data collection are expanding spectral coverage, while public datasets focus mainly on nutritional indices, lacking safety-related data. Framework-wide challenges persist in device compatibility, dataset comprehensiveness, and model interpretability. Recent advances show promising developments through: specialized deep learning architectures achieving 97-100% accuracy, data transformation techniques (2D-COS, GAFD) enhancing interpretability, hybrid traditional-deep learning models, and effective transfer learning for cross-device applications. Based on these insights, we propose three critical research directions: expanding food safety datasets through regulatory partnerships, developing multi-level fusion for heterogeneous device data, and creating automated techniques for model optimization and interpretability. These directions are vital for advancing ML-NIRS applications in food quality assessment, improving both efficiency and reliability.