Abstract
Accurate mapping of soil organic carbon (SOC) is essential for sustainable agricultural management, yet conventional stacking methods commonly use linear meta-learners, which provide a global combination of base model predictions but do not explain how the relative contributions of individual base models vary among samples. We propose AT-FC-Stacking, an interpretable ensemble framework integrating a sample-wise attention (AT) mechanism with a fully connected (FC) meta-learner to enable adaptive nonlinear fusion. Using Landsat 9 imagery, input variables were selected through combined Pearson correlation and Boruta feature selection, identifying key drivers such as Band 2 and sum average. AT-FC-Stacking substantially outperformed individual base models and linear benchmark models, achieving a mean R2 of 0.594, 10.41% higher than the traditional LR-Stacking model (mean R2 = 0.538), together with lower RMSE and MAE. The attention mechanism further revealed dynamic shifts in model contributions along the SOC gradient, with increasing contributions from extreme gradient boosting (XGBoost) and multilayer perceptron (MLP) but declining roles of support vector machine (SVM) and random forest (RF). This framework provides a clearer account of how the base learners contribute to the final prediction and thereby improves the interpretability of SOC estimation.