Satelliten-basierte Schadstoffkartierung: ML-Anwendungen und Analyse von atmosphärischen Einflussfaktoren

Abstract

The large-scale monitoring of air pollution is of central importance for public health and the effective implementation of measures to improve air quality. It is therefore essential to be able to distinguish which changes in air quality are caused by human activity and which can be attributed to weather phenomena. This work investigates the influence of meteorological factors on ground-level air pollution, as well as the spatial and temporal variability of its distribution. To this end, the concentrations of PM2.5, NO2 and O3 near the ground in an area containing Germany and parts of neighboring countries are derived using the Random Forest algorithm. It processes satellite observations, meteorological parameters and other auxiliary variables together with in-situ measurements of the pollutants. The trained Models provide reliable values for ground-level PM2.5 (R2 = 0.76) and NO2 (R2 = 0.79) and very good results for O3 (R2 = 0.91). Accuracy deviates when the test and training data are spatially or temporally separated. Methods of explainable artificial intelligence (XAI) are used to evaluate the significance of the individual model parameters. These show that the accuracy of the predictions for PM2.5 depends mainly on temporal information while NO2 relies on spatial and O3 on meteorological information. Heavy precipitation and strong winds are among the most important PM2.5- and NO2-reducing environmental factors, with air temperature and intense solar radiation greatly increasing O3-concentration. An investigation of the weekend effect on ozone production and the O3-NO2 model dependence suggest a shift from NOx-sensitive conditions in winter to NOx-saturated conditions in summer. Local Indicators of Spatial Association (LISA) are used to identify hot- and coldspots in the derived pollution data within the study area. Urban regions are hotspots for PM2.5 and NO2 while being coldspots for O3. Additionally, the surrounding geography and vegetation may play a role in the spatial distribution depending on the pollutant.