NDVI-Driven Dynamic Granger–Ramanathan Model by UAV Hyperspectral and Thermal Infrared Imagery for Leaf Water Content Estimation in Winter Wheat

Leaf water content (LWC) is a key indicator of winter wheat water status, but the reliability of UAV hyperspectral and thermal infrared predictions varies across samples, limiting the effectiveness of fixed-weight fusion. To improve LWC estimation by integrating hyperspectral and thermal infrared data, this study developed a dynamic Granger–Ramanathan averaging (GRA) method. Sensitive bands were selected using the successive projections algorithm, and random forest, XGBoost, and LightGBM models were constructed for each data source. Generalized additive models were used to characterize relationships between canopy–air temperature difference (ΔT), normalized difference vegetation index (NDVI), and single-source prediction errors, enabling dynamic adjustment of fusion weights. Random forest performed best among the single-source models, with test-set R2 values of 0.7871 and 0.6275 for hyperspectral and thermal infrared data, respectively. Conventional GRA achieved an R2 of 0.7645, whereas ΔT- and NDVI-driven dynamic GRA increased R2 to 0.8207 and 0.8224, respectively. The NDVI-driven model achieved the lowest RMSE (0.0646) and MAE (0.0497). Applied to 720 samples, it achieved an R2 of 0.8507, an RMSE of 0.0591, and an MAE of 0.0451. These findings demonstrate the potential of crop-status-driven dynamic weighting to improve multisource LWC estimation in winter wheat.