Scholarlink

<p>Urban PM₁₀ pollution under arid conditions is caused by intricate relations between human emissions and the chemistry in the atmosphere. Identification of pollutant drivers governing PM₁₀ variability remains challenging because many predictive models prioritize forecasting accuracy while offering limited insight into the relative influence of individual pollutants. Air quality observations of four monitoring stations in Riyadh (Al-Muruj, Al-Khaleej, Al-Khalidiya, and At-Taawun) were analyzed based on the combination of machine learning (ML) models with variance-based global sensitivity analysis to analyze station-specific pollutant effects towards PM₁₀ intensities. Three ML models&#xa0;with distinct learning mechanisms were employed, namely K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost). Model performance was evaluated against train–test splits (60/40–90/10) based on concordance correlation coefficient (DC) and normalized root mean square error (NRMSE). CatBoost exhibited the best generalization performance across the stations, with test DC values ranging from 0.86 to 0.89 and NRMSE values below 0.021 under moderate data partitioning schemes (60/40–70/30).&#xa0;Variance-based Sobol sensitivity analysis revealed significant spatial heterogeneity in pollutant influence. PM₁₀ variability at Al-Muruj (ST = 0.55 ± 0.04) was dominated by O₃, NO₂ at Al-Khaleej (ST = 0.61 ± 0.05) and CO at Al-Khalidiya (ST = 0.45 ± 0.04), and At-Taawun displayed mixed NO₂–O₃–SO₂ interactions. Integration of the ML with station-resolved global sensitivity analysis provides an interpretable framework for identifying pollutant contributions and interaction effects across monitoring locations.&#xa0;The findings&#xa0;provide a basis for location-specific air-quality assessment and targeted mitigation planning in urban environments.</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

Station-resolved Machine Learning and Variance-based Sensitivity Analysis of PM₁₀ Drivers in an Urban Environment

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Abdulhayat M. Jibrin,
  • Ali Aldrees,
  • Salisu Dan’azumi,
  • Sani I. Abba

摘要

Urban PM₁₀ pollution under arid conditions is caused by intricate relations between human emissions and the chemistry in the atmosphere. Identification of pollutant drivers governing PM₁₀ variability remains challenging because many predictive models prioritize forecasting accuracy while offering limited insight into the relative influence of individual pollutants. Air quality observations of four monitoring stations in Riyadh (Al-Muruj, Al-Khaleej, Al-Khalidiya, and At-Taawun) were analyzed based on the combination of machine learning (ML) models with variance-based global sensitivity analysis to analyze station-specific pollutant effects towards PM₁₀ intensities. Three ML models with distinct learning mechanisms were employed, namely K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost). Model performance was evaluated against train–test splits (60/40–90/10) based on concordance correlation coefficient (DC) and normalized root mean square error (NRMSE). CatBoost exhibited the best generalization performance across the stations, with test DC values ranging from 0.86 to 0.89 and NRMSE values below 0.021 under moderate data partitioning schemes (60/40–70/30). Variance-based Sobol sensitivity analysis revealed significant spatial heterogeneity in pollutant influence. PM₁₀ variability at Al-Muruj (ST = 0.55 ± 0.04) was dominated by O₃, NO₂ at Al-Khaleej (ST = 0.61 ± 0.05) and CO at Al-Khalidiya (ST = 0.45 ± 0.04), and At-Taawun displayed mixed NO₂–O₃–SO₂ interactions. Integration of the ML with station-resolved global sensitivity analysis provides an interpretable framework for identifying pollutant contributions and interaction effects across monitoring locations. The findings provide a basis for location-specific air-quality assessment and targeted mitigation planning in urban environments.