Application and Challenge of Machine Learning in the Satellite Remote Sensing of Atmospheric Aerosols

Abstract

Atmospheric aerosols play a crucial role in the atmospheric environment, impacting climate change and human health, garnering significant attention over the last fifty years. With the continuous development of satellite remote sensing technologies, satellite-based aerosol observation has become one of the most important and effective means of acquiring large-scale, long-term aerosol data. The advancement of satellite sensor technologies has also played a critical role in enabling highly accurate aerosol retrieval. Moreover, with the successful launch of POLDER (Polarization and Directionality of the Earth’s Reflectances), multi-angle joint polarization observation has become a key development direction for atmospheric aerosol monitoring. China has made significant progress in satellite-borne polarimeters, including DPC (Directional Polarimetric Camera) on GF-5 in 2018, SMAC (Synchronization Monitoring Atmospheric Corrector) on GFDM in 2020, PSAC (Polarized Scanning Atmospheric Corrector) on HJ-2A/B in 2020, and the PCF (polarization crossfire payload) on GF-5(02) in 2021. Subsequently, China launched a series of atmospheric environment monitoring satellites equipped with polarization crossfire sensors. Internationally, NASA launched the PACE satellite in 2024, equipped with SPEXone and HARP2 polarimeters. A review of the development of satellite remote sensing for atmospheric aerosols reveals that early observation methods were limited. Constrained by computational capacity and the few satellite data, early retrieval algorithms primarily relied on lookup table (LUT) approaches to strike a balance between retrieval accuracy and algorithmic feasibility. With advancements in satellite observation technologies, aerosol retrieval algorithms have continued to evolve. In the current era of massive satellite datasets, a central challenge in aerosol remote sensing is how to effectively extract meaningful atmospheric information from these vast data resources. Over the past decade, statistically optimized algorithms have emerged as one of the most promising solutions. These physically based inversion methods aim to retrieve aerosol parameters by fitting satellite observations with forward radiative transfer simulations. However, the high computational cost of optimization algorithms poses a significant barrier to the real-time generation of satellite aerosol products on a large scale, highlighting the urgent need for innovative technological breakthroughs. Recently, the evolution of Artificial Intelligence (AI) has introduced transformative changes to aerosol remote sensing. Machine learning (ML) techniques have shown notable enhancements in retrieval efficiency and the capability to tackle persistent issues that traditional physical methods face, such as separating signals from the surface and atmosphere. ML technologies are propelling satellite aerosol retrieval into an era of intelligent development. This paper presents a comprehensive review of the latest advancements in satellite-based aerosol retrieval using ML methods. It evaluates the strengths and limitations of mainstream ML techniques in different retrieval contexts. Overall, ML methods show great potential in aerosol remote sensing but still face limitations. It cannot yet fully replace physical models, but it offers promising prospects. Current research focuses on improving retrieval accuracy and efficiency for aerosol optical properties, while retrieval of microphysical and chemical properties remains exploratory. For near-surface PM concentrations, ML methods can build statistical models supported by meteorological data, yet underlying physical mechanisms need refinement. Integrating ML with physical models is a key future direction to enhance the accuracy and robustness of aerosol retrieval. This review aims to offer useful insights for researchers and developers working on next-generation aerosol retrieval systems.