Scholarlink

<p>Land surface temperature (LST) is a critical parameter in climate studies, influenced by meteorological and anthropogenic factors. This study integrates multiple satellite-derived products including MODIS land surface temperature (MOD11A1) and NDVI (MOD13A3) products with CHIRPS precipitation, TerraClimate meteorological variables, and Sentinel-2 land cover data, applying machine learning techniques to develop a comprehensive spatiotemporal framework for LST prediction in Khuzestan Province. Three regression models, Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multiple Linear Regression (MLR), were employed to predict LST using predictor variables including minimum and maximum air temperature, precipitation, evapotranspiration, soil moisture, vegetation index (NDVI), Digital Elevation Model (DEM), and land use/land cover (LULC) classes from 2001 to 2023. Feature importance analysis was conducted using model-derived importance scores and SHapley Additive exPlanations (SHAP) to quantify the contribution of each predictor. The analysis revealed that vegetation index (NDVI), air temperature, and elevation were the most influential variables governing LST variability. Among LULC-related variables, water cover showed moderate importance, whereas other land cover types had relatively minor effects. Quantile Regression Forest (QRF) was further used to quantify prediction uncertainty. Results demonstrated that RF and XGBoost significantly outperformed MLR, achieving RMSE = 1.5 and <i>R</i><sup>2</sup> = 0.92 compared to RMSE = 5.5 and <i>R</i><sup>2</sup> = 0.76 for MLR. Temporal predictions with XGBoost yielded RMSE values ranging from 1.4 to 6.6 and <i>R</i><sup>2</sup> from 0.68 to 0.98, highlighting its robustness for long-term modeling. Uncertainty analysis revealed that 55% of predictions fell within ± 2°C, 35% within ± 2 to ± 5°C, and 10% within ± 5 to ± 8°C, indicating reliable and interpretable results. This study underscores the effectiveness of RF and XGBoost in capturing complex LST dynamics and the value of SHAP analysis for identifying key drivers. Integrating uncertainty quantification enhances prediction reliability and provides a robust framework for climate research. The proposed approach is adaptable to other regions and offers a valuable tool for the monitoring of land-atmosphere interactions.</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

Machine learning insights into land surface temperature variability and prediction: a spatiotemporal approach with feature importance and uncertainty analysis

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Ali Rezaee,
  • Mohammad Reza Goodarzi,
  • Seyed Mohammad Alavizadeh,
  • Mojtaba Goldani

摘要

Land surface temperature (LST) is a critical parameter in climate studies, influenced by meteorological and anthropogenic factors. This study integrates multiple satellite-derived products including MODIS land surface temperature (MOD11A1) and NDVI (MOD13A3) products with CHIRPS precipitation, TerraClimate meteorological variables, and Sentinel-2 land cover data, applying machine learning techniques to develop a comprehensive spatiotemporal framework for LST prediction in Khuzestan Province. Three regression models, Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Multiple Linear Regression (MLR), were employed to predict LST using predictor variables including minimum and maximum air temperature, precipitation, evapotranspiration, soil moisture, vegetation index (NDVI), Digital Elevation Model (DEM), and land use/land cover (LULC) classes from 2001 to 2023. Feature importance analysis was conducted using model-derived importance scores and SHapley Additive exPlanations (SHAP) to quantify the contribution of each predictor. The analysis revealed that vegetation index (NDVI), air temperature, and elevation were the most influential variables governing LST variability. Among LULC-related variables, water cover showed moderate importance, whereas other land cover types had relatively minor effects. Quantile Regression Forest (QRF) was further used to quantify prediction uncertainty. Results demonstrated that RF and XGBoost significantly outperformed MLR, achieving RMSE = 1.5 and R2 = 0.92 compared to RMSE = 5.5 and R2 = 0.76 for MLR. Temporal predictions with XGBoost yielded RMSE values ranging from 1.4 to 6.6 and R2 from 0.68 to 0.98, highlighting its robustness for long-term modeling. Uncertainty analysis revealed that 55% of predictions fell within ± 2°C, 35% within ± 2 to ± 5°C, and 10% within ± 5 to ± 8°C, indicating reliable and interpretable results. This study underscores the effectiveness of RF and XGBoost in capturing complex LST dynamics and the value of SHAP analysis for identifying key drivers. Integrating uncertainty quantification enhances prediction reliability and provides a robust framework for climate research. The proposed approach is adaptable to other regions and offers a valuable tool for the monitoring of land-atmosphere interactions.