Physics-augmented hyperspectral sensing for soybean moisture assessment: Compensating packing density effects via computer vision

Physical variability during hyperspectral data acquisition can introduce measurement instability and compromise the reliability of moisture prediction in agricultural products. This study proposes a physics-augmented hyperspectral imaging (HSI) framework that explicitly quantifies and integrates key physical measurement factors into machine learning models for soybean moisture assessment. A total of 404 soybean seeds were conditioned to four moisture levels (6–17% w.b.) and used to systematically evaluate the effects of sample height, orientation, inter-sample distance (packing density), and light penetration depth on spectral responses. Optical analysis confirmed that soybean HSI reflectance is surface dominated, making spectral responses highly sensitive to packing configuration. Among the physical factors examined, inter-sample distance exhibited the strongest influence on spectral variability and was quantified via computer-vision-based distance measurement as an additional model input. The physical descriptor was incorporated into five predictive models (PLSR, RF, SVM, ANN, and XGB), which outperformed spectral-only baselines. The physics-augmented XGB model showed the highest performance (R 2 = 0.915, RMSE = 1.341) among the test models. SHAP analysis further demonstrated that the most influential wavelengths corresponded to O–H overtone and combination bands of water, suggesting the physical relevance of the learned spectral features. Overall, the results indicate that explicitly accounting for physical measurement factors enhances the robustness and interpretability of hyperspectral models and provides a practical strategy for improving measurement reliability in agricultural sensing applications. • Measurement-related physical factors govern hyperspectral moisture reliability. • Packing density introduces systematic spectral variability in soybean imaging. • XGB achieved highest external R 2 = 0.915 with interpretable SHAP analysis. • Results provide engineering guidance for hyperspectral system configuration.