• A dual-interpretation stacking method for base models and features was proposed. • Fusion of in-situ and satellite hyperspectral highlighted response to available HMs. • The stacking method has high explanatory for spatial variability of available HMs. • Adsorption and electron transition caused spectral response of available HMs. Machine learning (ML) models predicting the geospatial variability of soil heavy metals (HMs) often face monotonicity constraints and limited interpretability. Here, based on the satellite-borne hyperspectral combined with competitive adaptive reweighted sampling (CARS) and stepwise linear regression (SLR) models to reconstruct in-situ hyperspectral features, we proposed an interpretable stacking framework. In this framework, the Bayesian optimization was used to tune the hyperparameters of random forest (RF), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost), while partial least squares (PLS) served as the meta-model to predict available HMs (As, Cu, Ni, Zn) concentrations, and the SHapley additive explanations (SHAP) analysis quantified the prediction mechanisms of the stacking model and the contribution of characteristic bands. The results demonstrated that the stacking method significantly outperformed individual ML models (R 2 > 0.8). Spatial mapping of the stacking model indicated that Ni, Cu, and Zn exhibit relatively higher concentrations in the northern and eastern regions. Furthermore, the SHAP analysis indicated GBDT and XGBoost contributed the most (more than 30%) to predictions, followed by CatBoost (20%–30%) and RF (less than 20%). And the short-wave infrared band (1205–2500 nm) influenced As and Cu, while the visible light band (846–1088 nm) affected Ni and Zn. Moreover, the path analysis of the structural equation model (SEM) revealed that soil properties influence the spectral response mechanisms of available HMs by affecting the carbonate index (CAI) and soil organic matter (SOM). Overall, this study provides a dual interpretability stacking framework of both base models and hyperspectral features, offering new perspectives for enhancing productivity in agricultural remote sensing.
Pan et al. (Mon,) studied this question.