Atmospheric aerosols influence climate forcing, air quality, visibility, and human health, but their properties vary widely across space and time. Satellite instruments equipped with multi-angle and polarimetric sensors provide rich information about aerosol size, composition, and scattering behavior. However, traditional retrieval approaches—such as look-up tables or optimal estimation—struggle to fully exploit these high-dimensional observations and often require heavy computational resources. Deep learning methods offer efficiency but can lack physical interpretability and robustness when applied to complex atmospheric data. Based on these challenges, it is necessary to develop advanced retrieval strategies that can simultaneously integrate physical principles and data-driven learning to better characterize aerosols from modern satellite observations.
Researchers from the Aerospace Information Research Institute, Chinese Academy of Sciences, reported this work in Journal of Remote Sensing, published (DOI: 10.34133/remotesensing.1008) in December 2025. The study introduces a novel aerosol retrieval framework designed for data collected by the Directional Polarimetric Camera onboard the Gaofen-5(02) satellite. The new approach addresses a long-standing challenge in satellite remote sensing: how to efficiently and accurately extract aerosol optical depth, particle size information, and scattering properties from multi-angle, polarized measurements, especially during rapidly evolving haze and dust events.
The study demonstrates that a capsule network–based deep learning model can significantly improve aerosol retrieval accuracy from satellite data. By jointly processing multi-angle, multispectral, and polarization information, the model retrieves aerosol optical depth, fine-mode fraction, and single-scattering albedo with strong agreement to ground-based observations. Validation results show correlation coefficients exceeding 0.93 for aerosol optical depth and around 0.79 for particle size fraction, matching or surpassing the performance of widely used satellite products. Compared with conventional neural networks, the capsule-based approach better preserves the physical relationships among viewing geometry, polarization, and aerosol properties, enabling more stable retrievals across different pollution conditions.
The proposed framework integrates physical radiative transfer simulations with a capsule network architecture specifically designed to handle structured, multi-dimensional satellite observations. Simulated top-of-atmosphere reflectance and polarization signals were generated using a vector radiative transfer model under a wide range of aerosol and surface conditions. These simulations formed the training dataset for the deep learning model.
Unlike standard neural networks that compress information into scalar features, capsule networks represent features as vectors, allowing the model to encode both magnitude and directional relationships among observations. This design enables the algorithm to capture subtle variations in polarization and viewing angle that are critical for aerosol characterization. When applied to real satellite data over China in 2022, the model produced high-resolution aerosol maps consistent with ground-based measurements and existing satellite products. Case studies further showed that the method can clearly distinguish between fine-particle haze events and coarse dust storms, highlighting its capability for real-time air-pollution monitoring.
The research team noted that combining physical modeling with capsule networks offers a practical pathway to overcome the limitations of both traditional retrieval algorithms and purely data-driven approaches. They emphasized that the model's ability to process complex satellite observations efficiently makes it particularly suitable for operational monitoring, while its physical grounding improves robustness and interpretability for atmospheric science applications.
The study employed a hybrid methodology that couples vector radiative transfer simulations with deep learning. Synthetic satellite observations were generated across multiple wavelengths, viewing angles, and polarization states. These data were used to train a capsule network composed of primary and higher-level capsules with dynamic routing. The trained model was then applied to Gaofen-5(02) satellite observations, and retrieval results were validated using independent ground-based aerosol measurement networks to ensure accuracy and reliability.
The researchers envision that this approach could be rapidly adapted to other multi-angle polarimetric satellites, particularly during early mission phases when observational data are limited. Its near-real-time processing capability makes it valuable for tracking dust storms, haze episodes, and long-term aerosol trends. More broadly, the integration of physics-based modeling with advanced artificial intelligence could reshape how complex Earth-observation data are analyzed, supporting improved air-quality forecasting, climate assessment, and environmental decision-making worldwide.
###
References
DOI
10.34133/remotesensing.1008
Original Source URL
https://doi.org/10.34133/remotesensing.1008
Funding information
This work was supported by the National Outstanding Youth Foundation of China (Grant No. 41925019). Additional support was provided by the National Key R&D Program of China (Grant No. 2023YFB3907405) and the National Natural Science Foundation of China (Grant No. 42175148).
About Journal of Remote Sensing
The Journal of Remote Sensing, an online-only Open Access journal published in association with AIR-CAS, promotes the theory, science, and technology of remote sensing, as well as interdisciplinary research within earth and information science.