Domain-adapted modeling of wheat protein content across grain and flour morphological scales

Accurate and efficient estimation of grain protein content (GPC) plays a crucial role in wheat quality grading and processing quality assessment. Due to the differences in physical structure and optical scattering characteristics between grain and flour, models developed under a single morphology suffer from limited cross-morphology generalization performance, which makes it difficult to balance the integrated detection requirements for rapid quality evaluation and traceability. This study constructed a cross-morphology transfer learning framework to achieve bidirectional and high-precision collaborative prediction of GPC between grain and flour, thereby effectively overcoming the sample-morphology dependency of the models. This research investigated GPC prediction using hyperspectral data collected from both grain and flour samples. Original reflectance (OR) spectra and wavelet features (WF) were extracted to characterize spectral information. To evaluate feature relevance, Pearson correlation analysis, two-dimensional correlation spectroscopy (2D-COS), and variable importance in projection (VIP) analysis were applied. To address spectral discrepancies between grain and flour, a cross-morphology transfer-learning framework based on a wavelet-feature-enhanced Wasserstein distance adversarial neural network (WF-WDANN) was developed. Results showed that WF exhibited stronger correlations with GPC compared with OR features. For the proposed WDANN model, the highest prediction accuracies reached R² = 0.97 and NRMSE = 0.21% for grain, and R² = 0.99 and NRMSE = 0.19% for flour. In cross-morphology transfer scenarios, the WDANN model achieved an average independent validation performance of R² = 0.97, RMSE = 0.35%, and NRMSE = 0.17% for grain-to-flour transfer, and R² = 0.96, RMSE = 0.32%, and NRMSE = 0.15% for flour-to-grain transfer. These results demonstrate that integrating wavelet-based feature representation with adversarial transfer learning significantly improves GPC prediction across different sample morphologies, providing a robust and morphology-agnostic approach for rapid quality assessment in wheat processing chains.