Satellite-derived estimates of surface ozone by LESO: Extended application and performance evaluation

Abstract

Listen

Translate article

• Extending a surface O 3 estimation framework (LESO) from China to broader regions. • LESO captured surface O 3 spatiotemporal variations well in Europe and US. • Deep learning techniques improved the estimation performance by 30 %. • Satellite columns have relatively small impacts on estimating surface O 3 concentrations. • Spatially well-distributed in-situ sites are vital for global surface O 3 estimation. Surface O 3 pollution severely threats human health and crop production across the globe. Data-driven machine learning approaches are broadly used for estimating surface O 3 concentrations in China by integrating site-level measurements, meteorological reanalysis data, and satellite remote sensing observations. However, the feasibility of extending these approaches from China to the globe requires further investigation, and the sensitivities of different model drivers remain unknown. By using the LEarning Surface Ozone (LESO) framework, this study explores the effectiveness of estimating daily surface O 3 over Europe, the United States, and India and analyses the impacts of satellite columns, climate drivers, machine learning algorithms, and in-situ sites on the estimation performance. A cross-validation of the LESO framework against in-situ sites exhibited a promising performance with an mean R 2 higher than 0.9 and uncertainty smaller than 4 µg/m 3 in Europe and the United States. Due to its success in capturing the O 3 variability with a user-friendly interface and less computing resources occupancy, LESO is expected to promote efforts in global O 3 pollution monitoring. However, the estimation performance in India was not good enough (R 2 < 0.5 and uncertainty > 9 µg/m 3 ). Regarding potential factors affecting the estimation performance, the satellite O 3 columns and the number of in-situ sites did not affect the estimation performance significantly; the use of the deep learning technique improved the performance by 30 % than using a conventional non-deep learning algorithm; and the spatial sampling density of in-situ sites dominantly affected the estimation performance. This can be the reason causing the unsatisfactory performance in India. The findings of this study are beneficial to promoting global surface O 3 estimation and can bring evidence on the location selection for deploying new in-situ sites.