Satellite-based estimation of NO2 concentrations using a machine-learning model: A case study on Rio Grande do Sul, Brazil

Abstract

Nitrogen dioxide (NO₂) is one of the most important atmospheric pollutants, affecting human health (increasing susceptibility to respiratory infections) and the environment (soil and water acidification). In many regions of Brazil, NO₂ measurements at ground level meet difficulties because monitoring stations are few and unevenly distributed. Satellite observations combined with machine learning models can mitigate this lack of data. This paper report results from an investigation on the ability of a machine learning approach (a non-linear statistical Random Forest algorithm, hereafter RF) to reconstruct the long-term spatiotemporal ground-level NO₂ from 2006 to 2019 using as input parameters NO₂ data retrieved from the Ozone Monitoring Instrument (OMI) sensor aboard Aura satellite, besides meteorological covariates and localized ground-level NO₂ measurements. Results show that the RF model predicts NO₂ with an accuracy expressed by an R² = 0.68 correlation based on a 10-fold cross-validation. The model also predicted a mean NO₂ concentration of 18.73 ± 3.86 μg m^–3. The total NO₂ concentration over the entire region analyzed showed a decreasing trend (2.9 μg m^–3 yr^–1), being 2006 the year with the higher concentrations and 2017 with the lowest. This study suggests that non-linear statistical algorithm reconstructions using RF can be complementary tools to in situ and satellite observations for NO₂ mapping.