Enhanced continuous aerosol optical depth (AOD) estimation using geostationary satellite data: focusing on nighttime AOD over East Asia

TL;DR

This study develops machine learning models to estimate both daytime and nighttime aerosol optical depth over East Asia using geostationary satellite data, enabling continuous aerosol monitoring beyond daylight hours.

Contribution

The paper introduces novel machine learning models that utilize geostationary satellite data to estimate nighttime AOD, improving continuous aerosol monitoring capabilities.

Findings

The daytime TOA model achieved R2 of 0.83 and RMSE of 0.098.

Models showed high correlation with ground-based AERONET data.

SHAP analysis identified key variables influencing AOD estimation.

Abstract

Continuous aerosol monitoring in East Asia is essential due to the massive aerosol emissions from natural and anthropogenic sources. Geostationary satellites enable continuous aerosol monitoring; however, the observation is limited to the daytime. This study proposed machine learning-based models to estimate daytime and nighttime aerosol optical depth (AOD) in East Asia using a geostationary satellite, Geo-KOMPSAT-2A (GK-2A). The input variables for the machine learning models include the brightness temperature (BT) and top-of-atmosphere (TOA) reflectance from GK-2A, meteorological and geographical data, and auxiliary variables. The two models that used different combinations of GK-2A variables were proposed and compared: the all-day BT model, which estimates AOD during both day and night using BT variables, and the daytime TOA model, which estimates AOD during the day using TOA reflectance variables as well. The estimated AODs by the models were validated with ground-based AOD data from the Aerosol Robotic Network (AERONET) by 10-fold cross-validation and hold-out validation methods. The performance of the daytime TOA model was slightly higher than the all-day BT model during the day (R2 = 0.80-0.82, root mean square error (RMSE) = 0.107-0.116 for the all-day BT model, R2 = 0.83, RMSE = 0.098 for the daytime TOA model). The SHapley Additive exPlanations (SHAP) analysis showed that total precipitable water content and seasonality contributed the most for both proposed models. BT differences and TOA reflectance variables were identified as the next most contributing variables for the all-day BT and daytime TOA models. The spatiotemporal distributions of estimated AODs from the proposed models show similar patterns compared with other AOD products. A time series comparison at a test station demonstrated that the estimated AOD of the proposed models was consistent with the AERONET AOD.