Inversion of a Near‐Real‐Time China Gridded Hourly SO ₂ Emission Inventory Using Deep Learning Combined With 3D‐Variational Assimilation

Abstract Accurate emission inventories are essential for effective air quality prediction and management. Traditional bottom‐up methods and top‐down data assimilation approaches struggle to balance accuracy, timeliness, and usability. This study presents a novel inversion method combining deep learning (DL) and 3D‐Variational (3DVar) assimilation, termed DL‐3DVar. Based on the prior emissions and SO 2 concentration observations, this method was applied to invert China's grided hourly SO 2 emission inventory during the 2023 Chinese New Year (CNY). The results show that DL‐3DVar estimated an hourly SO 2 emission increment of 1.95 × 10 6 kg at midnight on CNY's Eve relative to the prior emissions, primarily due to extensive fireworks. Forecast experiments showed that using DL‐3DVar inverted emissions reduced the root‐mean‐square error (RMSE) by 52.9% and increased the correlation coefficient (CORR) by 331% compared to the prior emissions. This high‐temporal‐resolution emission inventory can improve pollutant forecasting and provide robust support for air pollution control policymaking.