Assessment of crude protein in wheat kernels using SWIR hyperspectral imaging combined with deep learning-based segmentation

Wheat crude protein content is a critical quality parameter for food processing, traditionally assessed using destructive chemical methods such as Kjeldahl or Dumas assays. While near-infrared (NIR) spectroscopy provides non-destructive analysis, its localized measurement approach limits throughput in large-scale applications. This study presents an integrated platform combining deep learning-based segmentation (YOLOv11) with hyperspectral imaging and Partial Least Squares Regression (PLSR) for rapid, non-destructive prediction of crude protein content in individual wheat kernels. We compared Visible and Near-Infrared (VNIR, 397–1004 nm) and Short-Wave Infrared (SWIR, 982–2577 nm) spectral regions for protein prediction. The YOLOv11 segmentation model achieved exceptional accuracy (F1 score >0.99) across both spectral ranges, enabling automated extraction of individual kernel spectra. Three preprocessing methods (SNV, MSC, and Savitzky-Golay) were systematically evaluated. The SWIR-based PLSR models consistently demonstrated superior generalization (all methods Q2 > 0.84), achieving a peak performance of Test R2 = 0.8725, RMSE = 0.7479, and Q2 = 0.8738 (using MSC). Notably, Savitzky-Golay preprocessing offered the optimal model simplicity, reducing PLS components from 10 to 3 while maintaining robust generalization (Test R2 = 0.8449, Q2 = 0.8487). In sharp contrast, the best VNIR model only achieved a Test R2 = 0.5750 (Q2 = 0.5793). YOLOv11 exhibited approximately 2.57 times faster inference speed compared to YOLOv8, substantially improving applicability for high-throughput analysis. The integration of high-speed deep learning segmentation with SWIR hyperspectral imaging establishes a practical framework for non-destructive quality assessment in grain processing facilities, quality control laboratories, and automated raw material sorting systems.