Scholarlink

<p>Dust storms and the associated air pollution pose significant environmental and health challenges in Inner Mongolia; however, accurately predicting PM10 concentrations remains difficult because of complex spatio-temporal patterns. This study provides a comprehensive comparison of machine learning and deep learning approaches for PM10 prediction in Inner Mongolia’s dust-dominated environment during the spring dust season (March–May, 2021–2024). All models achieved high accuracy (R² &gt; 0.90) but exhibited distinct strengths: LSTM excelled in temporal pattern recognition, CNN–LSTM in spatial feature extraction, XGBoost in computational efficiency with competitive accuracy, and ERT in capturing regional variations. We considered 15 candidate predictors; a genetic algorithm was then used to select an optimal subset (9 predictors) for the final model training. Bayesian optimization was used for hyperparameter tuning. The LSTM model achieved the highest overall accuracy (coefficient of determination = 0.93, root mean square error = 27.54&#xa0;µg/m³), followed closely by the CNN–LSTM model. For practical application, LSTM is recommended as the primary model for hourly PM10 prediction due to its best overall accuracy and stability. XGBoost provides a computationally efficient alternative with competitive skill under low-to-moderate concentrations, whereas ERT is more conservative for exceedance-risk screening but may overestimate high-risk areas. Spatial analysis revealed high pollution probabilities in southern and western Inner Mongolia, where more than 30% of the region has Level-1 PM10 pollution (50–150&#xa0;µg/m³) with probability greater than 0.6. Temporal analysis identified March as the peak month for dust pollution, with PM₁₀ concentrations occasionally exceeding 1200&#xa0;µg/m³ during severe dust events. Although all models tended to underestimate extreme concentrations, tree-based models showed relatively higher sensitivity to high-PM10 episodes.This study establishes a practical framework for dust-related PM10 prediction in arid regions and provides actionable insights for air-quality management and pollution-control strategies.</p> Graphical Abstract <p>This visual summary serves as a pivotal entry point into the research, offering an overview of the study’s core findings and methodologies. The left panel depicts PM₁₀ pollution over Inner Mongolia, highlighting the dust‐dominated environment and defining the regional context. Below, icons summarize the multi-source input data set: hourly PM₁₀ observations from the China National Environmental Monitoring Center for March–May 2021–2024, together with meteorological variables, vegetation indices, soil properties and topographic factors. The central workflow illustrates model development. Machine-learning models (ERT, XGB) are contrasted with deep-learning models (LSTM, CNN–LSTM). A genetic algorithm is used for feature selection, and Bayesian optimization is applied to tune hyperparameters. Model performance is evaluated using R², RMSE, MAE and MAPE. The panels on the right synthesize key results and implications. Boxplots indicate that deep-learning models outperform traditional approaches, with the LSTM model achieving the highest accuracy (R² = 0.93, RMSE = 27.54 µg m⁻³). Spatial probability maps reveal higher PM₁₀ pollution risks in southern and western Inner Mongolia, where around 30% of the area experiences Level-1 pollution (50–150 µg m⁻³). Seasonal boxplots emphasize March as the peak dust-pollution month, with PM₁₀ occasionally exceeding 1200 µg m⁻³. Overall, the graphical abstract highlights how deep-learning models effectively capture spatiotemporal PM₁₀ patterns, provide actionable insights into dust-related health risks and support regional air-quality management.</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 and Deep Learning of Dust Storm Pm10 in Inner Mongolia: Model Comparison and Spatial Pollution Risk

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Shunyu YAO,
  • Dongwei LIU,
  • Zhicheng QU

摘要

Dust storms and the associated air pollution pose significant environmental and health challenges in Inner Mongolia; however, accurately predicting PM10 concentrations remains difficult because of complex spatio-temporal patterns. This study provides a comprehensive comparison of machine learning and deep learning approaches for PM10 prediction in Inner Mongolia’s dust-dominated environment during the spring dust season (March–May, 2021–2024). All models achieved high accuracy (R² > 0.90) but exhibited distinct strengths: LSTM excelled in temporal pattern recognition, CNN–LSTM in spatial feature extraction, XGBoost in computational efficiency with competitive accuracy, and ERT in capturing regional variations. We considered 15 candidate predictors; a genetic algorithm was then used to select an optimal subset (9 predictors) for the final model training. Bayesian optimization was used for hyperparameter tuning. The LSTM model achieved the highest overall accuracy (coefficient of determination = 0.93, root mean square error = 27.54 µg/m³), followed closely by the CNN–LSTM model. For practical application, LSTM is recommended as the primary model for hourly PM10 prediction due to its best overall accuracy and stability. XGBoost provides a computationally efficient alternative with competitive skill under low-to-moderate concentrations, whereas ERT is more conservative for exceedance-risk screening but may overestimate high-risk areas. Spatial analysis revealed high pollution probabilities in southern and western Inner Mongolia, where more than 30% of the region has Level-1 PM10 pollution (50–150 µg/m³) with probability greater than 0.6. Temporal analysis identified March as the peak month for dust pollution, with PM₁₀ concentrations occasionally exceeding 1200 µg/m³ during severe dust events. Although all models tended to underestimate extreme concentrations, tree-based models showed relatively higher sensitivity to high-PM10 episodes.This study establishes a practical framework for dust-related PM10 prediction in arid regions and provides actionable insights for air-quality management and pollution-control strategies.

Graphical Abstract

This visual summary serves as a pivotal entry point into the research, offering an overview of the study’s core findings and methodologies. The left panel depicts PM₁₀ pollution over Inner Mongolia, highlighting the dust‐dominated environment and defining the regional context. Below, icons summarize the multi-source input data set: hourly PM₁₀ observations from the China National Environmental Monitoring Center for March–May 2021–2024, together with meteorological variables, vegetation indices, soil properties and topographic factors. The central workflow illustrates model development. Machine-learning models (ERT, XGB) are contrasted with deep-learning models (LSTM, CNN–LSTM). A genetic algorithm is used for feature selection, and Bayesian optimization is applied to tune hyperparameters. Model performance is evaluated using R², RMSE, MAE and MAPE. The panels on the right synthesize key results and implications. Boxplots indicate that deep-learning models outperform traditional approaches, with the LSTM model achieving the highest accuracy (R² = 0.93, RMSE = 27.54 µg m⁻³). Spatial probability maps reveal higher PM₁₀ pollution risks in southern and western Inner Mongolia, where around 30% of the area experiences Level-1 pollution (50–150 µg m⁻³). Seasonal boxplots emphasize March as the peak dust-pollution month, with PM₁₀ occasionally exceeding 1200 µg m⁻³. Overall, the graphical abstract highlights how deep-learning models effectively capture spatiotemporal PM₁₀ patterns, provide actionable insights into dust-related health risks and support regional air-quality management.