Using random forest to improve EMEP4PL model estimates of daily PM2.5 in Poland

Abstract

Long-term exposure to poor air quality is responsible for many diseases and increased mortality worldwide. European Environmental Agency reports that Poland is one of the most polluted countries in Europe due to high emissions associated with large coal and wood consumption and specific weather conditions. Exceedances of WHO-recommended PM2.5 thresholds are still common in Poland, so further action is needed to protect the health of the population. Atmospheric chemical transport models (CTMs) provide information on air quality to the public and are used to regulate air pollutant emissions. However, uncertainties associated with CTMs, related to e.g. physical/chemical processes and input data quality often lead to underestimation of pollutant concentrations, especially for PM2.5, and limits the applicability of CTMs e.g. for health impact studies.A hybrid approach combining the EMEP4PL chemical transport model for Poland with the Random Forest (RF) machine learning algorithm was applied to address the limitations of CTM and reduce its underestimation. We used EMEP4PL-modelled PM2.5 concentrations for period 2016–2019 as a predictor and measured daily PM2.5 from 71 monitoring stations in Poland as a dependent variable in three RF scenarios, which differed in terms of the selected predictors. The impact of different additional variables for the area was revealed, including population and emission data, dominant type of land use, Weather and Research Forecast (WRF) meteorological parameters, and temporal patterns of PM2.5 concentrations across the years. The models were evaluated in a random 5-fold and spatial leave-one-station-out cross-validations (LOSOCV), as well as for an independent test set. Our final model achieved a test set R2 of 0.71, compared to 0.38 for EMEP4PL, along with a reduction in negative bias (0.25 μg m−3 for the final RF, compared to −11 μg m−3 for EMEP4PL) and an improved ability to detect severe PM2.5 episodes. Enhanced coefficients of determination were observed in all seasons and at all monitoring sites included in the study, for both types of cross-validation and for the test set. We estimated the contribution of each group of variables separately and discovered, that the most impactful predictors for the area of Poland included temporal patterns of PM2.5 calculated based on the averages of EMEP4PL outcome (such as day of the year, week number, etc.) and meteorological modelled factors related to temperature, planetary boundary layer height, wind speed, and atmospheric pressure. The developed approach provides a basis for improved spatiotemporal estimates and forecasting of PM2.5 in the region, which is an important step toward better understanding the impact of air pollution on the local population well-being.