Land Surface Temperature (LST) downscaling is critical for analyzing urban thermal environments and mitigating Urban Heat Islands (UHIs), which exacerbate warming, energy demand, and public health risks. This study evaluates three machine learning (ML) models-Random Forest (RF), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost)-for downscaling Landsat-5 LST in Tabriz, Iran, a semi-arid and topographically complex city. Using 10-fold cross-validation, both tuned and untuned model configurations were assessed at 360 m and 600 m resolutions. RF (Tuned) achieved the highest accuracy (R-2 = 0.8163, RMSE = 1.18 degrees C at 360 m), outperforming Multiple Linear Regression (MLR), SVR, and XGBoost. At 600 m, SVR (Untuned) performed best (R-2 = 0.7955), and machine learning models demonstrated similar to 8 % less performance degradation than MLR under resolution coarsening. To improve model interpretability, Shapley Additive Explanations (SHAP) were employed, offering both global and local insights into feature importance. SHAP analysis identified NDBI and UI as dominant warming predictors in dense urban cores, while NDVI and MNDWI had stronger cooling effects in vegetated districts. These findings highlight the spatial heterogeneity of urban thermal drivers and support data-driven, equitable UHI mitigation strategies. Key innovations include the integration of hyperparameter tuning (yielding 2-5 % R-2 gains), SHAP-based explainability, and robust ML adaptability to heterogeneous urban landscapes. This approach directly supports climate-adaptive planning by delivering interpretable, high-resolution thermal data essential for sustainable urban design and environmental equity-key goals of resilient city development. The resulting LST maps provide actionable guidance for targeted interventions (e.g., green infrastructure, cool roofs, zoning reforms), offering transferable insights for climate-resilient development in Global South cities. By coupling hyperparameter tuning, SHAP-based explainability, and a systematic noise-trimming robustness analysis, our framework is computationally efficient, reproducible, and readily adoptable by planners in resource--constrained cities. Future work should incorporate 3D urban morphology, seasonal dynamics, and ensemble modeling to enhance decision support.