Estimation of surface ozone concentration over Jiangsu province using a high-performance deep learning model

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Data Science

Background:

• Global background ozone (O3) concentrations are rising, necessitating accurate monitoring.
• Ground-based O3 monitoring is reliable but requires high site density for spatial analysis.
• Existing machine learning models often fail to fully capture spatiotemporal O3 data.

Purpose of the Study:

• To develop a deep learning model for estimating daily maximum 8-hr average (MDA8) ozone concentrations.
• To capture the spatiotemporal characteristics of MDA8 O3 over Jiangsu province, China.
• To address limitations of existing models in providing comprehensive spatial and temporal O3 information.

Main Methods:

• Utilized a Residual connection Convolutional Long Short-Term Memory (R-ConvLSTM) network.
• Integrated TROPOMI total O3 column data, ERA5 meteorological data, and supplementary data for pre-training.
• Employed residual connections to mitigate gradient issues in deep networks.

Main Results:

• Achieved a high sample-based cross-validation R-squared of 0.955 and RMSE of 9.372 µg/m³.
• Demonstrated strong city-based cross-validation R-squared of 0.896 and RMSE of 14.029 µg/m³.
• Identified seasonal variations, with highest MDA8 O3 in spring (122.60 ± 31.60 µg/m³) and lowest in winter (69.93 ± 18.48 µg/m³).

Conclusions:

• The R-ConvLSTM model effectively estimates MDA8 O3 with high accuracy and spatiotemporal resolution.
• The model's architecture overcomes gradient issues, enabling deeper network analysis.
• Findings provide valuable insights into ozone pollution dynamics and inform environmental policy.