Spatially Explicit Ensemble Machine Learning for High-Resolution TVDI Mapping: Spatial Block Cross-Validation and SHAP Interpretability in a Topoclimatically Complex Mountainous Landscape

: Accurate spatial characterization of drought patterns across topographically complex terrain remains a critical challenge for adaptive land and water management. This study develops a robust spatial machine learning framework for continuous, high-resolution (30 m) mapping of the Temperature Vegetation Dryness Index (TVDI) across the A Luoi–Dakrong topoclimatic corridor in Central Vietnam — a landscape severely affected by Foehn-driven desiccation during the dry season. Three Bayesian-optimized ensemble algorithms — Random Forest, XGBoost, and CatBoost — were evaluated under standard accuracy metrics and a 5-fold spatial block cross-validation (SBCV) framework derived from empirical semi-variogram analysis. Under standard evaluation, all three models achieved comparable accuracy (R 2 = 0.8534, 0.8537, and 0.8576 for RF, XGBoost, and CatBoost, respectively); however, a directional divergence emerged under SBCV, with CatBoost demonstrating consistently superior spatial generalization (Spatial_CV_R 2 = 0.7783) relative to RF (0.7575) and XGBoost (0.7562) — a distinction that standard metrics alone cannot reveal. The framework identified pronounced drought heterogeneity across the corridor, with Dakrong district recording a substantially higher mean TVDI than A Luoi (0.524 vs. 0.388), and Krong Klang commune registering the most extreme conditions (mean TVDI = 0.834). SHAP-based interpretability analyses identified Elevation, NDWI, and Distance to Water as dominant drought controls, with drought stress declining monotonically from lowland (mean TVDI = 0.73, < 200 m) to montane zones (mean TVDI = 0.05, > 1200 m). Critically, spatially projected SHAP values uncovered pronounced bidirectional topographic non-stationarity: Elevation amplifies drought in the northern Dakrong valleys — where Foehn-driven adiabatic warming inverts expected lapse-rate cooling — while suppressing it across the southern A Luoi highlands, dismantling the assumption of a spatially universal elevation–drought relationship. The proposed framework constitutes a physically interpretable, high-resolution baseline tool for drought susceptibility mapping and spatially explicit risk prioritization in ecologically heterogeneous mountainous landscapes.