Evaluating four AOD Datasets as Predictors for PM2.5 using a Random Forest Approach

Kurzfassung

With a constantly growing population and urbanization health risks through air pollution become more and more severe. Especially fine particulate matter (PM2.5) is known to be one of the most dangerous pollutants causing serious diseases. A comprehensive monitoring of this pollutant is essential to better depict and understand its spatio-temporal variability and to assess its health effects. As station measurements provide only selective information, satellite observations of the aerosol optical depth (AOD) are widely used to monitor PM2.5 pollution with sufficient spatial coverage. In this study we apply a random forest approach to develop four PM2.5 prediction models for Germany, each incorporating a different AOD dataset together with a variety of additional proxy-data such as meteorological and land surface parameters. We evaluate the effect of resolution and coverage of the satellite data on the overall performance and accuracy of the random forest models. Moreover, we perform a feature analysis to investigate the influence of certain predictor variables on the model results. Beside the well-established AOD data from the Moderate resolution Imaging spectroradiometer (MODIS) onboard Terra and Aqua satellites, including Dark Target (DT) algorithm products and the Multi-Angle Implementation of Atmospheric Correction (MAIAC) product, we also employ more recent datasets from the Sea and Land Surface Temperature Radiometer (SLSTR) onboard Sentinel-3a and from the Tropospheric Monitoring Instrument (TROPOMI) operating on Sentinel-5 precursor. Overall, all models performed very well in predicting PM2.5. We could find an important dependency of the model performance on coverage and resolution of the AOD training data. Feature importance rankings show a higher weight of AOD as predictor variable compared to the proxy data, at least for two of the models.