Integrated hyperspectral–RGB modeling for the estimation of wheat protein and dough rheological properties using machine and deep transfer learning

The rapid non-destructive estimation of wheat protein properties and dough rheological characteristics is of great significance for ensuring processing quality, food safety, and variety classification. This study established a multi-source remote sensing framework integrating hyperspectral and red–green–blue (RGB) data. By extracting wavelet features (WFs) and color indices (CIs), and applying Pearson correlation analysis combined with the successive projections algorithm–variance inflation factor (SPA-VIF) for feature selection, partial least squares regression (PLSR), extreme gradient boosting (XGBoost), and attention-based multi-task learning (MTL-AM) were employed to predict wheat protein and rheological traits. The results showed that the fusion of wavelet features with RGB images significantly improved model prediction accuracy, and that the estimation performance of XGBoost and MTL-AM was significantly higher than that of PLSR. To further improve the estimation accuracy of the farinograph quality number (FQN), the combined MTL-AM–XGBoost model achieved the highest accuracy (R 2 = 0.982, RMSE = 8.975, MAE = 7.087, RPD = 6.077). This study provides an effective technical solution for the non-destructive, high-throughput monitoring of wheat functional quality.