A Machine Learning Framework for Retrieving Atmospheric Composition and the Associated Averaging Kernels From a Geostationary Hyperspectral Infrared Sounder

Abstract

Geostationary hyperspectral infrared sounders provide hourly observations of the atmospheric composition. However, the conventional optimal estimation method (OEM)‐based retrieval algorithm requires computationally expensive radiative transfer calculations and inverse modeling, making it difficult to monitor in real time. This study presents a machine learning (ML) framework that is trained using a small subset of OEM retrievals to retrieve both column density and associated averaging kernel (AK) information. This framework was applied to predict the columns of ammonia (NH 3 ) and formic acid (HCOOH) and the carbon monoxide (CO) profile over East Asia, using data from the Geostationary Interferometric Infrared Sounder (GIIRS) on the FengYun‐4B meteorological satellite. The results show that ML‐GIIRS reproduces OEM‐GIIRS retrievals with high accuracy ( R 2  > 0.9 for CO, R 2  > 0.8 for NH 3 and HCOOH), captures diurnal and seasonal variability and effectively recovers AK information, while achieving computation efficiency that is three orders of magnitude faster. However, prediction performance is slightly lower at night due to the reduced thermal contrast, which reduces detection sensitivity. Furthermore, the results demonstrate that the AK information, which quantifies the vertical sensitivity of the observing system, can be accurately predicted by the ML framework. Comparison with model simulations incorporating AK corrections shows that the ML‐GIIRS results exhibit negligible bias. Including AK prediction extends ML retrievals beyond column density estimates and supports the use of space‐borne hyperspectral infrared data for more physically interpretable applications. This developed ML framework has great potential to support near‐real‐time monitoring of air pollutants from geostationary hyperspectral infrared sounders.