English Abstract
Rapid and accurate soil organic carbon (SOC) quantification from UAV hyperspectral data is challenged by spectral redundancy and the limited generalization of individual models. This study proposes an RFE–TOPSIS–stacking method, a hybrid feature selection and stacking ensemble framework for SOC estimation using UAV hyperspectral data (350–1002 nm). Recursive feature elimination (RFE) based on XGBoost was first employed for preliminary band screening, retaining 19 bands. Then, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), a multi-criteria decision-making method, was applied to comprehensively evaluate these bands using mutual evaluation, integrating mutual information, entropy, random forest importance, the inverse coefficient of variation of reflectance, and band redundancy, and was then applied to select 12 SOC characteristic bands. A stacking regression model was subsequently constructed using a 1D-CNN, SVR, XGBoost, and PLSR as base learners and random forest as the meta-learner. The framework achieved an R2 of 0.85, an RMSE of 0.96 g/kg, and an RPD of 2.47 for the validation set, outperforming the individual model (1D-CNN: R2 = 0.78, RMSE = 1.05 g/kg, and RPD =2.26). Application to the three study areas revealed SOC contents (P10–P90) of 7.37–9.59, 7.52–11.01, and 9.98–15.43 g/kg; most validation sampling points had absolute residuals