Hybrid physics–AI aerosol property retrieval algorithm for AMI/GK-2A with a deep learning radiative transfer emulator

Minseok Kim; Jhoon Kim; Hyunkwang Lim; Seoyoung Lee; Hyeji Cha; Yujin Chai; Sang Seo Park; Jun Wang

doi:10.1016/j.jag.2026.105393

Back

Hybrid physics–AI aerosol property retrieval algorithm for AMI/GK-2A with a deep learning radiative transfer emulator

Journal article

Open access

Peer reviewed

Hybrid physics–AI aerosol property retrieval algorithm for AMI/GK-2A with a deep learning radiative transfer emulator

Minseok Kim, Jhoon Kim, Hyunkwang Lim, Seoyoung Lee, Hyeji Cha, Yujin Chai, Sang Seo Park and Jun Wang

International journal of applied earth observation and geoinformation, Vol.152, 105393

08/2026

DOI: 10.1016/j.jag.2026.105393

Files and links (1)

url

https://doi.org/10.1016/j.jag.2026.105393View

Published (Version of record) Open Access

Abstract

Computational advances have enabled the evolution and refinement of satellite aerosol retrieval algorithms to mitigate issues in previous versions and improve their results. This study addressed a fundamental limitation in operational satellite remote sensing which is the trade-off between computational efficiency and retrieval accuracy in conventional look-up table (LUT)-based algorithms. The hybrid-Yonsei Aerosol Retrieval (hybrid-YAER) algorithm for the Advanced Meteorological Imager (AMI) incorporates a flexible aerosol model input, which became available by using a deep learning (DL) radiative transfer model (RTM) that replaces the interpolation-based inversion with LUTs. The DLRTM, trained with 100,000 Vector LInearized Discrete Ordinate Radiative Transfer (VLIDORT) code simulations using a Set Transformer architecture, accurately reproduces top-of-atmosphere reflectance (the root-mean-square error is about 0.002 for Mie calculations) at computation times about 0.1% of those of a conventional RTM. Additionally, a dust call-back procedure and an extended surface reflectance database based on multi-year minimum reflectance are introduced in the hybrid-YAER algorithm. Validation against AERONET and comparison with Visible Infrared Imaging Radiometer Suite (VIIRS) Dark Target (DT) and Deep Blue (DB) AOD products demonstrate that the new algorithm mitigates underestimation of low values of aerosol optical depth (AOD) and discontinuities of AOD values between adjacent land and ocean, while improving overall correlation (R increased from 0.757 to 0.767 for land pixels and from 0.856 to 0.881 for ocean pixels) and reducing bias. Over ocean, bias patterns become more linear, and the retrieval stability is improved (R2 increased from 0.88 to 0.94) compared to the original version. Uncertainty analysis shows that retrieval uncertainty scales linearly with AOD, confirming the robustness of the hybrid algorithm. The concept of advancement in aerosol retrieval via hybrid algorithm can be applied to any other aerosol remote sensing algorithms.

Remote Sensing

Aerosol optical depth

Deep learning

Radiative transfer

Satellite

Details

Title: Subtitle: Hybrid physics–AI aerosol property retrieval algorithm for AMI/GK-2A with a deep learning radiative transfer emulator
Creators: Minseok Kim - Ulsan National Institute of Science and Technology
Jhoon Kim - Yonsei University
Hyunkwang Lim - National Institute for Environmental Studies (NIES), Tsukuba 305-0053, Japan
Seoyoung Lee - University of Maryland, Baltimore County
Hyeji Cha - Yonsei University
Yujin Chai - Yonsei University
Sang Seo Park - Ulsan National Institute of Science and Technology
Jun Wang - University of Iowa
Resource Type: Journal article
Publication Details: International journal of applied earth observation and geoinformation, Vol.152, 105393
DOI: 10.1016/j.jag.2026.105393
ISSN: 1569-8432
eISSN: 1872-826X
Publisher: Elsevier B.V
Grant note: Korean Ministry of Science and ICT (MSIT): N10250155 (00) National Research Foundation of Korea - MSIT: RS-2024-00346149 Korea Institute of Science and Technology: Global-24-004 Yonsei Fellow Program: Global-24-004
This work was supported by the InnoCORE program [grant number N10250155 (00) ] of the Korean Ministry of Science and ICT (MSIT), the National Research Foundation of Korea funded by MSIT [grant number RS-2024-00346149], and the Korea Institute of Science and Technology [grant number Global-24-004]. Jhoon Kim's work was supported by the Yonsei Fellow Program, funded by Lee Youn Jae. The authors gratefully acknowledge all principal investigators and their staff for establishing and maintaining the AERONET sites used in this study. The authors also acknowledge the NASA VIIRS aerosol teams for providing the Dark Target and Deep Blue products. The National Meteorological Satellite Center of the Korea Meteorological Administration is acknowledged for providing the GK-2A/AMI satellite data used in this research. We acknowledge helpful discussions with Dr. Edward J. Hyer.
Language: English
Date published: 08/2026
Academic Unit: Electrical and Computer Engineering; Civil and Environmental Engineering; Physics and Astronomy; Chemical and Biochemical Engineering
Record Identifier: 9985177837702771

Metrics

1 Record Views