Predicting Moisture Content in Sugar Beet Slices Using Shortwave Infrared Hyperspectral Imaging and Machine Learning

Abstract Accurate moisture monitoring in sugar beet ( Beta vulgaris L.) roots is critical for preserving sucrose, the main determinant of yield and economic value. Current industrial methods for measuring moisture are destructive and slow, limiting their suitability for real-time and large-scale applications. Previous studies have used hyperspectral imaging for sugar beet moisture prediction but relied mainly on visible to near-infrared wavelengths. No work has evaluated the stronger water absorption features available in the shortwave infrared region for sugar beet moisture assessment. This study addressed this gap by evaluating shortwave infrared (SWIR, 1000–2500 nm) hyperspectral imaging combined with machine learning for nondestructive prediction of root moisture content. Three commercial sugar beet varieties were grown under greenhouse conditions, and roots were sliced into 10-mm transverse sections ( n = 540). Spectra were acquired using a SWIR hyperspectral camera, and reference moisture content (56.9–88.4%) was determined gravimetrically after oven drying at 65 °C to constant weight. Three machine learning models, Artificial Neural Networks (ANN), Light Gradient Boosting Machine (LGBM), and One-Dimensional Convolutional Neural Networks (1D-CNN), were evaluated with three feature-reduction methods: principal component analysis (PCA), successive projections algorithm (SPA), and competitive adaptive reweighted sampling (CARS), under four preprocessing strategies: SNV, SG1, SNV + SG1, and no preprocessing. Across all models and feature-reduction strategies tested, ANN–CARS–SG1 provided the best overall performance (Rp = 0.808, RMSEP = 2.890%, RPDp = 1.663), while ANN–CARS–SNV + SG1 achieved the highest prediction correlation (Rp = 0.811). Among PCA-based models, 1D-CNN–PCA–SNV performed best (Rp = 0.781, RMSEP = 3.210%, RPDp = 1.497). Among SPA-based models, ANN–SPA–SG1 performed best (Rp = 0.773, RMSEP = 3.288%, RPDp = 1.461). Overall, CARS-based models provided the strongest predictive performance compared with PCA- and SPA-based configurations. These findings demonstrate the potential of SWIR hyperspectral imaging combined with wavelength selection and machine learning for rapid, nondestructive moisture assessment in sugar beet roots, while further validation is needed to improve quantitative prediction performance.