AI Summary & Key Takeaways
AI InsightThese results suggest that, in landscapes with pronounced seasonality, LST should be modeled separately for each season and feature importance should be interpreted on a seasonal basis.
📌 3 Key Takeaways
Abstract
This study evaluated three tree-based ensembles—Random Forest, XGBoost, and LightGBM—for seasonal modeling of land surface temperature (LST) in the Menteşe district of Muğla, Türkiye, a Mediterranean landscape with pronounced seasonality. Three nested feature sets—a topographic baseline extended first with hydrometeorological predictors and then with spectral indices—were compared for summer 2025 and winter 2025-2026, yielding 18 configurations across the three algorithms. Accuracy was assessed on a held-out random sample with bootstrap confidence intervals and Holm-corrected paired tests and feature attributions were quantified using Shapley additive explanations (SHAP) for each season. LightGBM with the full feature set was selected in both seasons (test R2 = 0.955, RMSE = 0.768 °C in summer; 0.932 and 0.525 °C in winter). Accuracy depended considerably more on the feature sets than on the algorithms: all feature-set increments remained significant after the Holm correction whereas the corrected tests detected no difference between LightGBM and XGBoost. Although all three feature groups contributed in each season, their relative contributions differed between seasons. The spectral indices dominated in summer, the topographic predictors held a nearly equal share in winter and the SHAP contribution of evapotranspiration reversed direction. The selected configurations were then applied across the district, and the resulting maps show a spatially heterogeneous summer pattern without a dominant gradient but a coherent winter gradient. These results suggest that, in landscapes with pronounced seasonality, LST should be modeled separately for each season and feature importance should be interpreted on a seasonal basis.