The study, published in the journal Engineering, was conducted through a collaborative effort by the Institute of Atmospheric Physics (IAP) of the Chinese Academy of Sciences, National Satellite Meteorological Center, the Harbin Institute of Technology, and other institutes.

 

Scientists believe that accurate measurements of atmospheric aerosols are pivotal in understanding Earth's radiation balance, climate change, and air quality. Aboard China's geostationary meteorological satellite FY-4A, the Advanced Geostationary Radiation Imager (AGRI) scans China every five minutes, providing crucial data for monitoring aerosol spatiotemporal variations.

 

However, the inflexibility of traditional physical retrieval algorithms, coupled with the insufficient number of ground-based sunphotometer sites, poses challenges in meeting the extensive sample requirements for machine learning in aerosol optical depth (AOD) retrieval.

 

In response to these challenges, the scientists developed an innovative AOD retrieval algorithm that combines deep learning and transfer learning. The new algorithm incorporates key concepts from the dark target and deep blue algorithms to facilitate feature selection for machine learning.

 

According to the study, independent validation confirms that the algorithm is highly accurate in estimating AGRI aerosol levels. The results show a strong correlation with expected values, indicating the algorithm's reliability in predicting aerosol optical depth.

 

"Our study showcases the significant potential of merging the physical approach with deep learning in geoscientific analysis," said the lead author Fu Disong from the IAP.

 

"The proposed algorithm holds promise for application to other multi-spectral sensors aboard geostationary satellites," Fu added.