Insertion-type near-infrared spectroscopy combined with interpretable machine learning for in situ prediction of oil content in stored soybeans

Accurate and rapid determination of oil content in stored soybeans is essential for quality control and economic evaluation, yet conventional chemical analysis is destructive, time-consuming, and impractical for large-scale in situ monitoring. This study proposes an insertion-type near-infrared (NIR) spectroscopy approach for non-destructive, real-time estimation of soybean oil content directly within bulk storage. Spectra were acquired using a self-developed probe-based NIR spectrometer and compared with a reference online device from the literature. Despite substantial differences in acquisition environments, both systems exhibited consistent spectral patterns, with key peaks near ∼1200 nm (first overtone of C–H stretching) and ∼1450 nm (combination of C–H stretching/bending and first overtone of O–H), indicating robustness across settings. Four modeling algorithms—Partial Least Squares Regression (PLSR), Gaussian Process Regression (GPR), Least Squares Support Vector Regression (LSSVR), and Kernel Extreme Learning Machine (KELM)—were tested with multiple preprocessing strategies. The optimal configuration, KELM with wavelet domain denoising, de-trending, and z-score standardization, achieved R² = 0.947, RMSE = 0.924, and RPD = 4.340, outperforming the other three models by at least 2.82 %, 6.38 %, and 20.16 %. SHAP analysis revealed the most influential wavelengths consistent with oil-related absorptions, demonstrating the feasibility of this in situ monitoring approach. • Developed insertion-type NIR system for in situ spectra inside bulk soybean storage. • Constructed dataset linking in situ NIR spectra with oil content in stored soybeans. • Built robust machine learning models to map NIR spectral features to oil content. • Applied SHAP to quantify band contributions and identify oil-related wavelengths.