Harnessing open remote sensing data and machine learning for daily ground-level ozone prediction models: Spatio-temporal insights in the continental biogeographical region

Abstract

Ground-level ozone (O₃) pollution poses significant environmental and public health challenges and requires accurate predictive models for effective monitoring and management. In this study we observe that 91 % of the observed ground-level O₃ variance can potentially be explained using time-lagged data from Sentinel-5P TROPOMI and data from ERA5-Land datasets on a trained artificial intelligence (AI) model deployed by machine learning (ML) in the continental part of the Veneto region in Italy. Data from local air quality monitoring stations were used as ground truth data. The study period is from January 2019 to December 2022. Spatio-temporal ML models predicted daily O₃ concentrations with RMSE of 9.05 μg/m³, 8.87 μg/m³ and 10.87 μg/m³ respectively for RF, XGB and LSTM. Models without spatio-temporal information gave lower accuracy, with RMSE of 10.88 μg/m³, 11.45 μg/m³ and 12.06 μg/m³ respectively, showing that spatio-temporal information can improve performance more than 10 %. However, spatio-temporal independent models are more transferable across continental region and different seasons. Results provide spatially continuous maps of ground-level O₃ with a spatial resolution of ∼11.13 km (0.1°), helping to estimate pollution levels in areas without ground stations. Spatial analysis of the models’ performance showed consistent high accuracy across all stations, while temporal analysis revealed lower performance in summer months. Overall, while the spatial resolution of the models developed in this study is insufficient for risk management in urban areas, they have practical implications for daily ground-level O₃ monitoring in areas without ground stations in the continental region.