Ml and Geostatistical Methods for Spatial Downscaling Air Quality Data

Air quality monitoring processes aimed at preventing environmental risks at small areas, require very high spatial resolution information for several hazardous air pollutants, as well as for those meteorological conditions that most influence air quality. However, almost all the available data sets usually refer to long time-series recorded irregularly over the area of interest, at a coarse spatial resolution. In this context, effective methods for the spatial downscaling are needed to interpolate/forecast the variable of interest in a higher resolution, both in univariate and multivariate scenario. In this paper, a mixed approach based on blind source separation, deep kriging and graph neural networks is proposed for spatial downscaling and an application to ozone concentrations and meteorological variables is developed.