Method Adaptation Patterns and a Selection Framework for Near-Infrared Spectroscopy-Based Moisture Prediction of Major Grain Crops

Abstract

Moisture content is an important indicator of grain storage safety, processing quality, and circulation efficiency. However, the performance of near-infrared spectroscopy (NIRS) models is affected by differences in grain crops, sample morphologies, preprocessing strategies, feature selection methods, and regression models, and the adaptation relationships among these factors remain insufficiently understood. In this study, soybean, maize, and wheat samples with whole-grain and powder morphologies were investigated to reveal the method adaptation patterns of NIRS-based moisture prediction under consistent experimental conditions. Spectra in the range of 900–1700 nm were collected, and 54 analytical pathways were established by combining six preprocessing strategies, three feature selection algorithms (SPA, CARS, and UVE), and three regression models (PLSR, SVR, and RF). The results showed that representative optimal pathways achieved Rp values above 0.9730 for powder samples and above 0.9587 for whole-grain samples. Different grain crops and sample morphologies exhibited distinct analytical pathway preferences, indicating that appropriate analytical strategies should be selected according to specific detection objects. Spectral variation analysis further demonstrated higher spectral variability in whole-grain samples than in powder samples. This study provides insights into the selection of suitable NIRS analytical strategies for grain moisture prediction and quality assessment.