Scholarlink

<p>Urban heat islands are intensifying in rapidly urbanizing megacities, with major implications for public health, infrastructure, and environmental equity. Understanding surface urban heat island (SUHI) patterns is challenging because land surface temperature (LST) responds to nonlinear processes and spatially heterogeneous urban form. Here, we develop an explainable Geospatial Artificial Intelligence (GeoAI) framework to model summer (May–August 2020) LST across Dhaka, Bangladesh, using a 500&#xa0;m grid and multi-source predictors describing urban morphology and land cover. An AutoML pipeline (FLAML) selected a LightGBM regressor, achieving strong predictive accuracy (R<sup>2</sup> = 0.725, RMSE = 1.18&#xa0;°C, MAE = 0.83&#xa0;°C). To interpret model behavior, we compare SHAP with GeoShapley, which decomposes predictions into non-spatial feature contributions, an intrinsic location effect, and location–feature interaction terms that capture spatially varying relationships. Results indicate that built-up intensity is the dominant predictor of higher LST, while vegetation and water are associated with lower LST, with substantial geographic variation in the magnitude and direction of these relationships. GeoShapley decomposition highlights localized cooling thresholds for green and blue infrastructure and identifies spatial clusters of intrinsic thermal drivers not explained by observed variables. Overall, the proposed framework provides spatially explicit, policy-relevant interpretation for targeting heat mitigation, and is transferable to other data-constrained cities.</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

Understanding urban heat patterns in Dhaka city using explainable GeoAI and satellite-derived LST

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Maria Sultana,
  • Md Didarul Islam,
  • Bradford Johnson

摘要

Urban heat islands are intensifying in rapidly urbanizing megacities, with major implications for public health, infrastructure, and environmental equity. Understanding surface urban heat island (SUHI) patterns is challenging because land surface temperature (LST) responds to nonlinear processes and spatially heterogeneous urban form. Here, we develop an explainable Geospatial Artificial Intelligence (GeoAI) framework to model summer (May–August 2020) LST across Dhaka, Bangladesh, using a 500 m grid and multi-source predictors describing urban morphology and land cover. An AutoML pipeline (FLAML) selected a LightGBM regressor, achieving strong predictive accuracy (R2 = 0.725, RMSE = 1.18 °C, MAE = 0.83 °C). To interpret model behavior, we compare SHAP with GeoShapley, which decomposes predictions into non-spatial feature contributions, an intrinsic location effect, and location–feature interaction terms that capture spatially varying relationships. Results indicate that built-up intensity is the dominant predictor of higher LST, while vegetation and water are associated with lower LST, with substantial geographic variation in the magnitude and direction of these relationships. GeoShapley decomposition highlights localized cooling thresholds for green and blue infrastructure and identifies spatial clusters of intrinsic thermal drivers not explained by observed variables. Overall, the proposed framework provides spatially explicit, policy-relevant interpretation for targeting heat mitigation, and is transferable to other data-constrained cities.