Knowledge based convolutional transformer for joint estimation of PM_2.5 and O₃ concentrations.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Chemistry

Background:

• Traditional air pollutant concentration prediction methods focus on single pollutants, requiring significant time and effort.
• Accurate prediction of multiple air pollutants is crucial for environmental monitoring and public health.

Purpose of the Study:

• To develop an efficient and accurate model for the collaborative estimation of PM2.5 and O3 concentrations.
• To incorporate domain knowledge into a data-driven approach for improved air quality prediction.

Main Methods:

• Proposed a Convolutional Transformer (Convtrans) model integrating ground, satellite, and reanalysis data.
• Introduced knowledge through shared/specific inputs, a PM2.5-O3 interaction module, and a weighted loss function.
• Compared Convtrans with CNN-LSTM, Transformer, RF, and XGB models for validation.

Main Results:

• Convtrans demonstrated minimum error (PM2.5 RMSE: 6.136 µg/m³, O3 RMSE: 8.250 µg/m³) and highest accuracy (PM2.5 R²: 0.923, O3 R²: 0.898) in typical Chinese cities.
• Generated spatial variation maps of pollutant concentrations based on model predictions.
• Validated the feasibility of integrating knowledge into data-driven models for joint pollutant estimation.

Conclusions:

• The Convtrans model offers a feasible and effective approach for the joint estimation of atmospheric pollutant concentrations.
• The proposed joint estimation framework has potential applications in multivariate retrieval across various scientific fields.