Enhanced soil moisture estimation: A hybrid framework integrating UAV hyperspectral data and dielectric models

Accurate field-scale estimation of soil moisture content (SMC) in red-soil farmland supports hydrological monitoring and agricultural water management. This study assessed whether soil dielectric constant (SDC) information could improve unmanned aerial vehicle (UAV) hyperspectral estimation of SMC and provide a dielectric-guided reference for predictions. Hyperspectral imagery (400–1000 nm) and field measurements of SMC and SDC were collected at 125 locations. Spectral-index baselines, dielectric models, hyperspectral-only machine-learning models, data-driven models incorporating measured SDC, and dielectric-constrained hybrid models were evaluated using spatially blocked five-fold cross-validation; feature selection and hyperparameter optimization were conducted within each training fold. Reflectance near the 970-nm water-absorption region was repeatedly identified as informative. Local calibration markedly improved the dielectric relationship: the modified Topp model achieved a mean R 2 of 0.9621, an RMSE of 0.0114 cm 3 / cm 3 , and an RPD of 5.5497. Among hyperspectral-only models, random forest performed best ( R 2 = 0.6259; RMSE = 0.0349 cm 3 / cm 3 ). Incorporating measured SDC produced the highest numerical accuracy in the data-driven random forest model ( R 2 = 0.9786; RMSE = 0.0076 cm 3 / cm 3 ), although Shapley additive explanations (SHAP) analysis indicated that SDC dominated this improvement. Applying the locally calibrated dielectric relationship as a prediction boundary improved all three hyperspectral models. Hybrid models achieved mean R 2 values of 0.8633–0.8904 and RMSE values of 0.0191–0.0218 cm 3 / cm 3 , with Light Gradient Boosting Machine (LightGBM) performing best. Study demonstrate that dielectric information can complement hyperspectral observations and provide a useful methodological reference for improving field-scale SMC estimation.