Scholarlink

<p>Outdoor thermal stress is intensifying across Türkiye, but its meteorological controls remain poorly resolved across contrasting climate regimes. We use explainable machine learning to decompose model-attributed contributions to daily maximum Universal Thermal Climate Index (UTCI) variability, rather than to independently predict a derived thermal index. ERA5 and ERA5-HEAT data for 1950–2025 were used to assemble 826,799 warm-season grid-day observations with UTCI &gt; 26&#xa0;°C at 0.25° × 0.25° resolution. A Random Forest model was fitted using seven predictors: 2-m air temperature (T2M), relative humidity, vapor pressure deficit (VPD), wind speed, surface solar radiation downwards, surface thermal radiation downwards (STRD), and specific humidity. Model performance was treated as an internal consistency check for attribution (R² = 0.936, RMSE = 1.80&#xa0;°C). Spatial and temporal blocked validation confirmed stable SHAP rankings, with T2M and VPD ranked first and second in all 12 folds (Kendall’s W = 0.963). Humidity-correlation, perturbation-mode, and ablation tests showed that VPD represents a robust moisture-demand dimension rather than an independent humidity effect. Severity-stratified SHAP showed that VPD attribution share increased from 18.7% under moderate heat stress to 28.7% under very strong heat stress, while wind-speed and STRD shares declined. Pixel-wise SHAP identified VPD dominance across the semi-arid interior, wind sensitivity along the Black Sea and Marmara margins, and localized STRD dominance along parts of the Mediterranean coast. Very strong heat grid-days increased significantly (+ 5.0 grid-days yr⁻¹), whereas extreme heat showed a non-significant increase (+ 0.62 grid-days yr⁻¹). These regimes can inform regionally differentiated heat-risk adaptation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spatial and severity-dependent controls of thermal comfort across Türkiye: a SHAP-based decomposition of UTCI drivers

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Doğukan Doğu Yavaşlı

摘要

Outdoor thermal stress is intensifying across Türkiye, but its meteorological controls remain poorly resolved across contrasting climate regimes. We use explainable machine learning to decompose model-attributed contributions to daily maximum Universal Thermal Climate Index (UTCI) variability, rather than to independently predict a derived thermal index. ERA5 and ERA5-HEAT data for 1950–2025 were used to assemble 826,799 warm-season grid-day observations with UTCI > 26 °C at 0.25° × 0.25° resolution. A Random Forest model was fitted using seven predictors: 2-m air temperature (T2M), relative humidity, vapor pressure deficit (VPD), wind speed, surface solar radiation downwards, surface thermal radiation downwards (STRD), and specific humidity. Model performance was treated as an internal consistency check for attribution (R² = 0.936, RMSE = 1.80 °C). Spatial and temporal blocked validation confirmed stable SHAP rankings, with T2M and VPD ranked first and second in all 12 folds (Kendall’s W = 0.963). Humidity-correlation, perturbation-mode, and ablation tests showed that VPD represents a robust moisture-demand dimension rather than an independent humidity effect. Severity-stratified SHAP showed that VPD attribution share increased from 18.7% under moderate heat stress to 28.7% under very strong heat stress, while wind-speed and STRD shares declined. Pixel-wise SHAP identified VPD dominance across the semi-arid interior, wind sensitivity along the Black Sea and Marmara margins, and localized STRD dominance along parts of the Mediterranean coast. Very strong heat grid-days increased significantly (+ 5.0 grid-days yr⁻¹), whereas extreme heat showed a non-significant increase (+ 0.62 grid-days yr⁻¹). These regimes can inform regionally differentiated heat-risk adaptation.