New AI-Powered Algorithm Enhances Satellite Monitoring of Air Pollution

A research team led by Prof. SHI Chong from the Aerospace Information Research Institute (AIR) of the Chinese Academy of Sciences (CAS), in collaboration with Japanese scientists, has made a breakthrough in satellite-based air quality monitoring. The team has developed AIRTrans, an innovative algorithm that dramatically improves the accuracy and efficiency of retrieving key aerosol properties from multispectral satellite observations.

Aerosols—tiny particles suspended in the atmosphere—play a major role in air pollution, human health, and climate change. Two important indicators for studying aerosols are Aerosol Optical Thickness (AOT), which measures the amount of particles in the air, and Fine-Mode Fraction (FMF), which indicates the proportion of small particles such as those from vehicle emissions or industrial processes.

Traditional satellite retrieval algorithms based on pre-calculated lookup tables utilize only two or three spectral channels, which limits the detail and accuracy of the results. Full-physics optimal estimation methods offer greater precision by leveraging more spectral information, but they are often too slow for real-time applications and struggle with complex aerosol types.

To address these challenges, the team developed AIRTrans, a hybrid algorithm that combines physical modelling with machine learning method. The algorithm uses a radiative transfer model RSTAR to simulate satellite spectral reflectance with diverse aerosol types and surface properties, generating a rich database to pre-train a neural network model. The model is then fine-tuned with ground-based measurements, improving its accuracy and adaptability to various real-world scenarios. The algorithm also accounts for hourly changes in surface reflectance and background AOT, further refining its accuracy.

AIRTrans is specially designed to work with multispectral instruments like Advanced Himawari Imager (AHI) onboard Himawari-8, which observes the East Asia region every 10 minutes. This allows for real-time, high-frequency tracking of aerosol changes throughout the day—something traditional methods struggle to achieve.

When tested against ground-based measurements, AIRTrans achieved a root mean square errors of just 0.132 for AOT and 0.146 for FMF—a significant improvement about 40% and 49%, respectively, compared to the official Himawari-8/AHI products.

Notably, AIRTrans resolved the systematic underestimation of FMF in existing retrieval products. During the real-world events like dust storms and haze events, AIRTrans successfully captured the spatiotemporal dynamics of aerosol concentrations and particle sizes, demonstrating its value for pollution monitoring and early warning.

The team’s work shows that AIRTrans is not only accurate and fast but also widely applicable. It can be adapted for use with other multispectral satellite sensors, providing powerful support for precise air pollution control and climate change, especially in the context of carbon neutrality goals.

The research was published in the Remote Sensing of Environment  under the title "Development of a hybrid algorithm for the simultaneous retrieval of aerosol optical thickness and fine-mode fraction from multispectral satellite observation combining radiative transfer and transfer learning approaches". The project was supported by the National Natural Science Foundation of China.

True-color images and haze probability maps depict a Feb 12, 2021 North China haze event. AIRTrans data show decreasing aerosol optical thickness (AOT) from 03:00–08:00 UTC, aligned with lower haze probability, while elevated fine-mode fraction (FMF) indicates anthropogenic aerosols. (Image by AIR)