Improved Himawari-8 10-minute scale aerosol optical depth product using deep neural network over Japan

Abstract

Aerosol Optical Depth (AOD) is critical to air quality research and has a major impact on the Earth's energy budget. Recently, satellite retrieval has been a common data source for large-scale AOD information, for which the new-generation geostationary Himawari-8 satellite, which monitors East Asia and the Pacific region every 10 min, provides an unprecedented opportunity to track its dynamics. However, in the previous evaluation research, the Himawari-8 L2 AOD product has a large deviation from the ground observation. In this study, we proposed a deep neural network (DNN) model to correct the existing Himawari-8 L2 AOD using the synchronous information recorded in the Himawari-8 spectral bands in Japan. The results show that the accuracy has been dramatically improved, as indicated by the random 10-fold cross-validation (CV) (R2: 0.76 vs. 0.24; and RMSE: 0.068 vs. 0.194). Furthermore, many more samples fall in the error lines of the Global Climate Observation System (improved from 24.7% to 52.0% in 10-fold CV). In addition, we also obtained better performance when evaluated using the leave-one-station-out CV (R2 = 0.49 and RMSE = 0.101). The Shapley Additive exPlanations values (SHAP) based feature importance result showed that the Day of Year (DOY) contributed the most to the AOD retrievals, followed by the Himawari-8 L2 AOD. The performance decreased when either DOY or Himawari-8 L2 AOD was removed from the model. Overall, the results of this study will provide a methodology to obtain the high temporal resolution AOD with improved accuracy and thus will help to quantify the dynamics of aerosols in a synchronous manner.