Joint Retrieval of PM2.5 Concentration and Aerosol Optical Depth over China Using Multi-Task Learning on FY-4A AGRI

Abstract

Aerosol information, particularly PM2.5 concentration and aerosol optical depth (AOD), is essential for inquiries pertaining to air quality, human health impacts, and climate change. Previous approaches to estimate PM2.5 concentration from remote sensing data have largely relied on AOD retrieval techniques. However, the complexity of the AOD–PM2.5 interplay and limitations of existing algorithms pose significant challenges to achieving precise co-estimations of the two quantities at the same location. On this point, a multi-task learning (MTL) model, which enables the joint retrieval of PM2.5 concentration and AOD, is proposed and applied on the top-of-the-atmosphere reflectance data gathered by the Fengyun-4A Advanced Geosynchronous Radiation Imager (FY-4A AGRI). Using the Multi-Angle Atmospheric Correction (MAIAC) AOD and ground-based PM2.5 observations as references, the performance of the proposed MTL model is evaluated and compared to that of two single-task learning (STL) models, namely, random forest (RF) and deep neural network (DNN). Specifically, the MTL model estimates that the coefficient of determination (R2) of AOD is 0.88 and the root mean squared error (RMSE) is 0.10, while the estimated R2 of PM2.5 is 0.84, and the RMSE is 13.76 μg·m–3. Compared with the RF model, the R2 of the MTL model for estimating AOD and PM2.5 increased by 0.04 and 0.06, respectively, and the RMSE decreased by 0.02 and 4.55 μg·m–3, respectively. In addition, compared with the DNN model, the R2 of the MTL model for estimating AOD and PM2.5 increased by 0.01 and 0.05, respectively, and the RMSE decreased by 0.02 and 1.76 μg·m–3, respectively. The evaluation suggests that the MTL model is able to provide simultaneously improved AOD and PM2.5 retrievals, offering a critical advantage in capturing the actual distribution of fine particulate matter with high time efficiency, and thereby substantially advancing satellite retrieval techniques for air quality research and climate change studies.