Remote Sensing Fine Estimation Model of PM_2.5 Concentration Based on Improved Long Short-Term Memory Network: A Case Study on Beijing–Tianjin–Hebei Urban Agglomeration in China

Abstract

The accurate prediction of PM2.5 concentration across extensive temporal and spatial scales is essential for air pollution control and safeguarding public health. To address the challenges of the uneven coverage and limited number of traditional PM2.5 ground monitoring networks, the low inversion accuracy of PM2.5 concentration, and the incomplete understanding of its spatiotemporal dynamics, this study proposes a refined PM2.5 concentration estimation model, Bi-LSTM-SA, integrating multi-source remote sensing data. First, utilizing multi-source remote sensing data, such as MODIS aerosol optical depth (AOD) products, meteorological data, and PM2.5 monitoring sites, AERONET AOD was used to validate the accuracy of the MODIS AOD data. Variables including temperature (TEMP), relative humidity (RH), surface pressure (SP), wind speed (WS), and total precipitation (PRE) were selected, followed by the application of the variance inflation factor (VIF) and Pearson’s correlation coefficient (R) for variable screening. Second, to effectively capture temporal dependencies and emphasize key features, an improved Long Short-Term Memory Network (LSTM) model, Bi-LSTM-SA, was constructed by combining a bidirectional LSTM (Bi-LSTM) model with a self-adaptive attention mechanism (SA). This model was evaluated through ablation and comparative experiments using three cross-validation methods: sample-based, temporal, and spatial. The effectiveness of this method was demonstrated on Beijing–Tianjin–Hebei urban agglomeration, achieving a coefficient of determination (R2) of 0.89, root mean squared error (RMSE) of 12.76 μg/m3, and mean absolute error (MAE) of 8.27 μg/m3. Finally, this model was applied to predict PM2.5 concentration on Beijing–Tianjin–Hebei urban agglomeration in 2023, revealing the characteristics of its spatiotemporal evolution. Additionally, the results indicated that this model performs exceptionally well in hourly PM2.5 concentration forecasting and can be used for PM2.5 concentration hourly prediction tasks. This study provides technical support for the large-scale, accurate remote sensing inversion of PM2.5 concentration and offers fundamental insights for regional atmospheric environmental protection.

Keywords

• MODIS AOD
• self-adaptive attention mechanism
• long and short-term memory network
• deep learning
• PM_2.5 concentration prediction