LESO: A ten-year ensemble of satellite-derived intercontinental hourly surface ozone concentrations

Abstract

This study presents a novel ensemble of surface ozone (O3) generated by the LEarning Surface Ozone (LESO) framework. The aim of this study is to investigate the spatial and temporal variation of surface O3. The LESO ensemble provides unique and accurate hourly (daily/monthly/yearly as needed) O3 surface concentrations on a fine spatial resolution of 0.1◦ × 0.1◦ across China, Europe, and the United States over a period of 10 years (2012–2021). The LESO ensemble was generated by establishing the relationship between surface O3 and satellite-derived O3 total columns together with high-resolution meteorological reanalysis data. This breakthrough overcomes the challenge of retrieving O3 in the lower atmosphere from satellite signals. A comprehensive validation indicated that the LESO datasets explained approximately 80% of the hourly variability of O3, with a root mean squared error of 19.63 μg/m3. The datasets convincingly captured the diurnal cycles, weekend effects, seasonality, and interannual variability, which can be valuable for research and applications related to atmospheric and climate sciences.

DeepSAT4D: Deep learning empowers four-dimensional atmospheric chemical concentration and emission retrieval from satellite

“…[58][59] The horizonal domain covers most of East Asia and is represented by grids of 182 rows and 232 columns with a horizontal resolution of 27 km by 27 km. The whole troposphere (from ground to 100mb) is defined by using 14 layers with sigma values as follows: 1.00, 0.995, 0.99, 0.98, 0.96, 0.94, 0.91, 0.86, 0.8, 0.74, 0.65, 0.55, 0.4, 0.2, and 0.00, corresponding to 19,57,114,230,386,584,910,1375,1908,2618,3598, 5061, 7620, 11944 meters above the ground at domain and annual averaged level.…”

“…[8][9] Nevertheless, persistent challenges remain in the accurate estimation of ground concentrations from satellite retrievals. These challenges stem from difficulties in assigning column density to vertical layer-specific concentrations, whether through the application of statistic regressions, [10][11] machine learning techniques [12][13][14][15][16][17][18][19][20] or numerical model simulations. [21][22][23] The primary obstacle lies in the lack of accurate representation of vertical profiles, which exhibit strong spatial and temporal variations.…”