A kriging-calibrated machine learning method for estimating daily ground-level NO₂ in mainland China

Area of Science:

• Environmental Science
• Atmospheric Chemistry
• Geostatistics

Background:

• Accurate estimation of daily nitrogen dioxide (NO2) levels using combined satellite and ground monitoring data remains a challenge.
• Conventional cross-validation (CV) provides average model performance, masking significant grid-to-grid variations in prediction accuracy.
• Evaluating model performance across different spatial grids and identifying factors influencing extrapolation ability are crucial but underexamined.

Purpose of the Study:

• To compare the daily NO2 estimation capabilities of three distinct methods across mainland China (2014-2016).
• To develop a novel two-stage meta-analysis method for assessing the impact of ground site number and distribution on grid-level prediction accuracy.
• To identify key factors influencing the extrapolating ability of NO2 estimation models.

Main Methods:

• Developed and compared three NO2 estimation methods: universal kriging, a satellite-based approach (Non-linear exposure-lag-response & Extreme gradient boosting), and a kriging-calibrated satellite method.
• Employed a novel two-stage meta-analysis to explore the influence of nearby site distribution and quantity on prediction accuracy.
• Utilized cross-validation (CV) to evaluate model performance, focusing on R-squared and root mean square error (RMSE).

Main Results:

• The kriging-calibrated satellite method achieved superior performance with a CV R-squared of 0.85 and RMSE of 7.87 μg/m³, outperforming the universal kriging (CV R²=0.57) and satellite-based methods (CV R²=0.81).
• The two-stage meta-analysis demonstrated that model performance degrades with sparser ground site distribution.
• Adding five ground sites within a 50 km radius improved model extrapolating ability by 17.51%.

Conclusions:

• Kriging calibration enhances the accuracy of satellite-based machine learning models for NO2 prediction.
• The developed meta-analysis method provides effective guidance for strategically placing new ground monitoring sites to maximize budget-limited improvements.
• Optimizing ground site networks is essential for improving the reliability of satellite-derived NO2 estimations.