Deep learning-driven reconstruction of PM2.5 vertical profiles: A fusion of lidar and tower data

Abstract

Understanding the vertical distribution of fine particulate matter (PM_2.5) is essential for developing effective air pollution management and control strategies. To thoroughly investigate and capture the spatio-temporal evolution patterns of PM_2.5, this study has developed a deep-learning-based reconstruction algorithm for estimating its vertical concentration distribution, leveraging the fusion of lidar observations and tower-based monitoring experiments. The results reveal that although multiple machine learning techniques can effectively estimate the concentration of PM_2.5, the Convolutional Neural Network and Bidirectional Long Short-Term Memory (CNN-BiLSTM) model exhibited the best estimation effect, achieving an R² value of 0.98. Building upon this model, the study further discusses the seasonal and daily variation characteristics of PM_2.5. The vertical distribution analysis shows that the high PM_2.5 concentrations are primarily clustered below 1000 m. Further composite analysis illustrates a significant positive correlation between PM_2.5 and ozone (O₃), another significant pollutant in composite pollution, at altitudes ranging from 200 to 3000 m, likely due to shared sources and mutual interactions. However, specific seasonal and altitudinal patterns emerge, such as a negative correlation observed between 1200 and 1800 m in autumn and 800 and 1200 m in winter. This phenomenon may be attributed to a threshold effect, indicating that low concentrations of PM_2.5 do not necessarily induce significant changes in O₃ concentration. This study provides crucial insights into the vertical monitoring for PM_2.5 and offers new perspectives on the complex interaction mechanism between PM_2.5 and O₃ within the context of composite pollution.