Estimating daily PM2.5 concentrations using an extreme gradient boosting model based on VIIRS aerosol products over southeastern Europe


Gündoğdu S., Tuna Tuygun G., Li Z., Wei J., Elbir T.

AIR QUALITY ATMOSPHERE AND HEALTH, vol.15, no.12, pp.2185-2198, 2022 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Volume: 15 Issue: 12
  • Publication Date: 2022
  • Doi Number: 10.1007/s11869-022-01245-5
  • Journal Name: AIR QUALITY ATMOSPHERE AND HEALTH
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, IBZ Online, ABI/INFORM, BIOSIS, CAB Abstracts, Geobase, Pollution Abstracts, Veterinary Science Database
  • Page Numbers: pp.2185-2198
  • Keywords: PM2.5 estimation, XGBoost, Aerosol optical depth, VIIRS, Southeastern Europe, EMERGENCY-ROOM VISITS, LONG-RANGE TRANSPORT, RANDOM FOREST MODEL, PARTICULATE MATTER, AIR-POLLUTION, METEOROLOGICAL FACTORS, PM10 CONCENTRATIONS, SATELLITE DATA, OPTICAL DEPTH, ONE DECADE
  • Dokuz Eylül University Affiliated: Yes

Abstract

The performance of aerosol optical depth (AOD) products from the visible infrared imaging radiometer suite (VIIRS) instrument to estimate ground-level PM2.5 concentrations has been determined at different locations; however, it is still limited over Europe. VIIRS dark target (DT) and deep blue (DB) AOD products at 6-km spatial resolution and independent variables from the MERRA-2 reanalysis were used for estimating daily PM2.5 concentrations in southeastern Europe. An estimation model based on the Extreme Gradient Boosting (XGBoost) approach was developed and tested for DT and DB AODs. The estimations were compared with daily PM2.5 observations from 122 air quality monitoring stations in five countries, including Bulgaria, Cyprus, Greece, Romania, and Turkey. The estimated PM2.5 concentrations were consistent with ground measurements with the Pearson correlation coefficient (R) of 0.82 and 0.78, showing overall low estimation uncertainties with the root mean square error (RMSE) of 7.43 and 8.38 mu g/m(3) and the mean absolute error (MAE) of 4.76 and 5.31 mu g/m(3) for DT and DB AOD datasets, respectively. Independent model results were also discussed based on each country and season. The best estimation accuracy reached the R value of 0.83 with an average RMSE of 9.05 mu g/m(3) and an MAE of 5.84 mu g/m(3) in Turkey with DB AOD. In contrast, the model with DT AOD was highly accurate with the R value of 0.85, showing minor overall uncertainties (i.e., RMSE = 6.08 and 3.31 mu g/m(3)) over Greece. The highest accuracies were obtained in autumn and spring, while the lowest ones were available in winter and summer. This study provides a feasible machine learning approach to estimate PM2.5 using VIIRS AOD products in southeastern Europe.