Related Experiment Video
Updated: Dec 7, 2025
Collecting and Processing Drone-based Remotely Sensed Data for Use in Forest Recovery Monitoring
Published on: October 24, 2025
Ensemble-based deep learning for estimating PM2.5 over California with multisource big data including wildfire smoke.
Lianfa Li1, Mariam Girguis2, Frederick Lurmann3
1Department of Preventive Medicine, University of Southern California, Los Angeles, CA, USA; State Key Laboratory of Resources and Environmental Information System, Institute of Geographical Sciences and Natural Resources, Chinese Academy of Sciences, Beijing, China.
This study developed an advanced deep learning model to accurately predict fine particulate matter (PM2.5) concentrations and their uncertainties across California. The model effectively captures complex air quality dynamics, aiding crucial health effect studies.
Area of Science:
- Environmental Science
- Data Science
- Atmospheric Science
Background:
- Accurate estimation of fine particulate matter (PM2.5) concentrations and prediction uncertainties is vital for public health studies.
- California presents unique challenges for PM2.5 prediction due to diverse natural and anthropogenic emissions, complex meteorology, topography, and land use patterns.
Purpose of the Study:
- To develop a high spatiotemporal resolution PM2.5 prediction model with uncertainty estimates for California.
- To leverage ensemble-based deep learning and big data fusion for improved air quality modeling.
Main Methods:
- Utilized autoencoder-based full residual deep networks and ensemble learning to model complex PM2.5 interrelationships.
- Integrated diverse data sources including remote sensing (MAIAC AOD, NDVI), MERRA-2, wildfire smoke data (HYSPLIT), meteorology, land cover, and traffic data.
- Imputed missing MAIAC AOD observations and calculated wildfire smoke contribution to PM2.5.
Main Results:
- Achieved high prediction accuracy with a test RMSE of 2.29 μg/m³ (R²: 0.87) for PM2.5.
- Identified key predictors including M2GMI CO, temporal basis functions, spatial location, air temperature, MAIAC AOD, and PM2.5 sea salt mass.
- Model demonstrated strong performance on independent AQS site data and accurately captured wildfire-related PM2.5 spatial and temporal patterns.
Conclusions:
- The developed ensemble deep learning method offers a generalizable approach for PM2.5 prediction in complex environments.
- Prediction uncertainty estimates provide valuable insights for exposure assessment and health studies.
- The model's ability to handle diverse environmental factors enhances its applicability in regions with challenging air quality prediction needs.
