Application of machine learning (individual vs stacking) models on MERRA-2 data to predict surface PM_2.5...

Area of Science:

• Environmental Science
• Atmospheric Science
• Data Science

Background:

• Spatial monitoring of PM2.5 in India is uneven due to limited ground stations.
• Modern-Era Retrospective Analysis for Research and Applications Version 2 (MERRA-2) provides estimates but shows significant underestimation compared to ground observations.
• Machine learning (ML) offers a potential solution to improve PM2.5 estimation accuracy.

Purpose of the Study:

• To assess the utility of ML models in enhancing PM2.5 estimations using MERRA-2 data.
• To compare individual ML models and a stacking technique for improved PM2.5 prediction.
• To evaluate the performance of the best ML model across different regions in India.

Main Methods:

• MERRA-2 aerosol and meteorological data were used as input features.
• Individual ML models (XGBoost, Random Forest, LightGBM) were trained and tested.
• A stacking technique with XGBoost as the meta-regressor was employed.
• Model performance was validated against ground observations across India.

Main Results:

• XGBoost (XGB) demonstrated the best individual model performance (r² = 0.73).
• The stacking technique, using XGB as a meta-regressor, outperformed individual models (r² = 0.77).
• The stacked model provided good hourly PM2.5 predictions across India, with the eastern region showing the best results (r² = 0.80).
• Performance varied by region, with the central region showing the lowest accuracy (r² = 0.46) due to sparse data.

Conclusions:

• ML, particularly the stacking technique, effectively improves PM2.5 estimation from MERRA-2 data in India.
• The developed model offers enhanced hourly PM2.5 predictions, outperforming daily predictions.
• Regional variations in performance highlight the impact of data availability on ML model accuracy.