Grain geometry matters: Hyperspectral imaging challenges for moisture detection in individual kernels of barley

Abstract

This study evaluated the effects of penetration depth, grain orientation, and placement angle on hyperspectral imaging (HSI) signal quality using short-wave infrared (SWIR) HSI (1000–2500 nm). Orientation-related spectral variability was observed, primarily due to groove direction and angular placement. Penetration assessment with lead sulfide (PbS) quantum dots (QDs) at 1200 nm and 1800 nm revealed that barley husks exhibited higher transmittance, while intact kernels showed limited light penetration. Using Extreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) analyses, two key moisture sensitive wavelength regions (1153 nm and 1954 nm) were identified. Deep learning models, including 2D and 3D convolutional neural network (CNN), were developed and evaluated under several configurations. The 3D CNN showed the best performance when trained solely on HSI data with selected feature wavelengths. Using 540 barley kernels for training and 132 kernels for testing, the model achieved an R² of 0.98, an RMSE of 1.1763, and an MAE of 1.2592, demonstrating that spectral–spatial information alone can provide stable and accurate moisture prediction, and that further inclusion of geometric metadata (angle and orientation) provides limited benefit. The proposed XGBoost–SHAP–3D CNN framework offers an interpretable, efficient, and cost-effective solution for rapid moisture estimation in barley, demonstrating strong potential for intelligent grain quality monitoring.

Plant phenotyping relevance

大麦個粒の水分という植物器官の状態を、HSIとXGBoost–SHAP–3D CNNで推定する手法を開発・評価しており、表現型取得・推定手法が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.