Scholarlink

<p>Against the backdrop of accelerating economic growth and urbanization, air pollution—particularly PM<sub>2.5</sub>—remains a global public health challenge. While China has made significant progress in PM<sub>2.5</sub> reduction since 2015, regional disparities and seasonal variations persist, necessitating refined spatiotemporal modeling. This study introduces an innovative LSTM-XGBoost hybrid framework to overcome limitations of conventional models. Leveraging PM<sub>2.5</sub> ground monitoring, satellite-derived AOD, meteorological datasets, and DEM data (2015–2020), we systematically addressed temporal inconsistencies (e.g., daily/monthly predictors) through temporal aggregation and spatial disaggregation. The workflow entailed data preprocessing, where we harmonized 1&#xa0;km × 1&#xa0;km resolution datasets and imputed missing values using temporal interpolation; feature engineering, involving extracting spatial features (e.g., land use, elevation) via XGBoost and temporal dependencies (e.g., seasonal patterns) via LSTM; model integration, which fused XGBoost’s spatial pattern recognition with LSTM’s sequential learning to capture nonlinear spatiotemporal interactions; and validation, employing 10-fold cross-validation and independent testing against RF, XGBR, and standalone LSTM. The results of the study show that (1) according to the ranking results of the model performance evaluation index R<sup>2</sup>, the LSTM-XGBoost model exhibits the highest accuracy, followed by the RF, XGBR and LSTM models. (2) By linearly fitting the measured and estimated values in different years and seasons, it is found that the R<sup>2</sup> is higher than 0.94 and the R<sup>2</sup> is highest in winter, followed by autumn, spring, and summer, which verifies the model’s high efficiency and stability in the estimation of PM<sub>2.5</sub> concentration; (3) The PM<sub>2.5</sub> concentration from 2015 to 2020 shows a gradual decrease in the trend of change, and the monthly average value shows a ‘U’-shaped characteristics, showing obvious seasonal fluctuation characteristics, and at the same time, it shows that the concentration in winter, autumn, spring and summer decreases sequentially; (4) Through spatial analysis, it is found that the PM<sub>2.5</sub> concentration in the seasonal scale shows the spatial distribution characteristics of high in the northern cities and low in the southern 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

Estimation of PM2.5 concentration and its temporal and spatial distribution using LSTM-XGBoost modeling

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Aoxin Zhao,
  • Zhenxue Zou,
  • Gang Cheng,
  • Huan Ding

摘要

Against the backdrop of accelerating economic growth and urbanization, air pollution—particularly PM2.5—remains a global public health challenge. While China has made significant progress in PM2.5 reduction since 2015, regional disparities and seasonal variations persist, necessitating refined spatiotemporal modeling. This study introduces an innovative LSTM-XGBoost hybrid framework to overcome limitations of conventional models. Leveraging PM2.5 ground monitoring, satellite-derived AOD, meteorological datasets, and DEM data (2015–2020), we systematically addressed temporal inconsistencies (e.g., daily/monthly predictors) through temporal aggregation and spatial disaggregation. The workflow entailed data preprocessing, where we harmonized 1 km × 1 km resolution datasets and imputed missing values using temporal interpolation; feature engineering, involving extracting spatial features (e.g., land use, elevation) via XGBoost and temporal dependencies (e.g., seasonal patterns) via LSTM; model integration, which fused XGBoost’s spatial pattern recognition with LSTM’s sequential learning to capture nonlinear spatiotemporal interactions; and validation, employing 10-fold cross-validation and independent testing against RF, XGBR, and standalone LSTM. The results of the study show that (1) according to the ranking results of the model performance evaluation index R2, the LSTM-XGBoost model exhibits the highest accuracy, followed by the RF, XGBR and LSTM models. (2) By linearly fitting the measured and estimated values in different years and seasons, it is found that the R2 is higher than 0.94 and the R2 is highest in winter, followed by autumn, spring, and summer, which verifies the model’s high efficiency and stability in the estimation of PM2.5 concentration; (3) The PM2.5 concentration from 2015 to 2020 shows a gradual decrease in the trend of change, and the monthly average value shows a ‘U’-shaped characteristics, showing obvious seasonal fluctuation characteristics, and at the same time, it shows that the concentration in winter, autumn, spring and summer decreases sequentially; (4) Through spatial analysis, it is found that the PM2.5 concentration in the seasonal scale shows the spatial distribution characteristics of high in the northern cities and low in the southern cities.