Spatial downscaling of CAMS surface pollutants using Machine Learning

In order to monitor regional pollution over Europe, the Copernicus Atmosphere Monitoring Service (CAMS) coordinated by the European Centre for Medium-Range Weather Forecasts (ECMWF) implemented an operational multi-model air quality forecast system over Europe. CAMS regional products provide a 5-day forecast of several chemical species (e.g. NO2, CO, PM2.5, PM10) from the surface up to 5 km with a spatial resolution of 10 km. In addition, CAMS global services provide similar products globally in a coarser resolution of 0.4° (40 km approximately).

Motivated by the fact that air pollution is a global problem and responsible for millions of premature deaths each year, combined with the lack of a 10 km global air quality forecast system, we train a convolutional neural network (CNN) in order to downscale the spatial resolution of CAMS surface NO2 from 40 km to 10 km. Since most pollutants are affected by meteorological conditions and topographic characteristics, we use as input several meteorological variables (e.g. wind, temperature, humidity, boundary layer height) from ECWMF high-resolution forecasts (HRES) as well as surface elevation and emission information of several pollutants. All inputs are available at 10 km resolution globally.

In order to validate if there is an added value in our downscaled results, we evaluate against observations collected by a network of surface stations. Our downscaling efforts in this study focus over the European domain, where the reference of a high-resolution chemistry is available from the CAMS regional services, but we aim to train a model that will be general enough for global application.