A machine learning framework for urban heat mitigation in informal settlements: Climate-resilient planning in Kabul, Afghanistan

Rising temperatures disproportionately impact vulnerable communities in informal settlements of arid Global South cities, yet data-driven frameworks for heat-resilient planning remain limited. This study pioneers an integrated machine learning (ML) framework—combining multivariate clustering, ensemble models (Random Forest, XGBoost, Gradient Boosting), and SHAP explainability—to analyze Land Surface Temperature (LST) dynamics in Kabul, Afghanistan. Results reveal informal settlements endure significantly higher LST (up to +5°C) than formal areas, driven by dense low-rise structures, minimal green space, and adjacent barren lands. While Gradient Boosting achieved the highest predictive accuracy (R² ≈ 0.45), the core contribution lies in translating ML insights into actionable planning strategies derived from urban morphological indicators (UMIs): (1) an optimal vegetation threshold (NDVI ≈0.15), (2) building heights around 3m to balance shade and ventilation, and (3) vertical densification for population management. Seasonal analysis highlights adaptive planning needs, with UMIs exerting stronger influences in summer but remaining relevant year-round. This research provides a replicable methodology for UMI-LST analysis in informal settlements, offering a pathway for equitable, climate-resilient urban development. We urge policymakers to embed targeted greening, managed densification, and land-use optimization into Kabul’s urban agenda.