Researchers from the Hefei Institutes of Physical Science of the Chinese Academy of Sciences, have developed an intelligent algorithm to improve aerosol optical depth (AOD) retrieval from China’s spaceborne polarimetric remote sensing data. The new method, based on an attention-enhanced Kolmogorov–Arnold network, improves the accuracy and reliability of aerosol monitoring using observations from the High-Precision Polarization Scanner (POSP) onboard the Gaofen-5B satellite.
The findings were published in IEEE Transactions on Geoscience and Remote Sensing.
AOD is an important measure of atmospheric aerosols, with applications in air quality monitoring and climate research. However, existing retrieval methods face challenges under complex conditions and often make limited use of satellite spectral and polarization information.
In this study, researchers developed an attention-enhanced Kolmogorov–Arnold network (AKAN) to improve AOD retrieval from satellite observations. The model combines the learning ability of Kolmogorov–Arnold networks with an attention mechanism, helping it focus on important information from multispectral polarimetric data.
The team tested AKAN using observations from POSP, a high-precision polarization sensor independently developed by the researchers. By matching satellite data with ground-based measurements from the Aerosol Robotic Network, they built a dataset of more than 240,000 samples for model training and testing.
AKAN achieved accurate global AOD retrieval, with an R2 value of 0.9336 and most results meeting the expected accuracy requirements.
The researchers also used the SHapley Additive exPlanations method to examine how the model reached its predictions. The analysis showed that the important spectral bands and scattering-angle information selected by the model agreed well with known atmospheric processes, indicating that the AI model can provide both accurate results and useful information for understanding aerosol observations.
Further tests under different scenarios showed that the method maintained good accuracy and adaptability under various atmospheric and surface conditions.
The new method offers a promising approach for improving aerosol monitoring with satellite remote sensing data.
Global AOD distribution maps. Global AOD distribution in (a) February, (b) May, (c) August, and (d) November. (Image by HUANG Honglian)
