2025-04-09 中国科学院(CAS)
＜関連情報＞
- https://english.cas.cn/newsroom/research_news/infotech/202504/t20250410_1040901.shtml
- https://www.sciencedirect.com/science/article/pii/S0034425725000239
マルチスペクトル衛星観測からのエアロゾル光学的厚さとファインモード分率の同時検索のための放射伝達と伝達学習アプローチを組み合わせたハイブリッドアルゴリズムの開発 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
Chenqian Tang, Chong Shi, Husi Letu, Shuai Yin, Teruyuki Nakajima, Miho Sekiguchi, Jian Xu, Mengjie Zhao, Run Ma, Wenwu Wang
Remote Sensing of Environment Available online: 11 February 2025
DOI:https://doi.org/10.1016/j.rse.2025.114619
Highlights
- Novel hybrid algorithm combining radiative transfer and transfer learning for aerosol retrieval.
- Improved simultaneous retrieval accuracy for AOT and FMF from Himawari-8/AHI.
- Dynamic aerosol model configuration covering multiple aerosol types.
- Captures aerosol spatial and temporal variations in haze and dust events.
Abstract
The aerosol optical thickness (AOT) and fine-mode fraction (FMF) are crucial to understanding the radiative and environmental effects of aerosols. However, accurately retrieving these properties simultaneously from monodirectional multispectral satellite data remains challenging. Inversion algorithms based on lookup tables typically leverage information from only two or three channels, resulting in limited retrieval parameters. Although optimal estimation methods can enhance the utilization of multispectral information, they are mostly constrained by fixed aerosol types and have higher computational overhead due to the multiple iterations. To achieve real-time, high-precision, and simultaneous retrieval of the AOT and FMF for geostationary satellites with high-frequency observation, we propose a novel hybrid algorithm, AIRTrans, for the Himawari-8/AHI by integrating radiative transfer (RT) and transfer learning (TL) approaches. Specifically, RT is used to construct a simulation dataset that covers multiple aerosol types and surface conditions corresponding to the simulated multispectral observation, which pre-trains an artificial neural network model. The TL strategy is then employed to fine-tune this model using in situ data, enhancing its representativeness in real scenarios. AIRTrans performs direct retrieval using satellite observations and surface reflectance constructed via the second minimum reflectance method but considering background AOT. Results indicate that the AIRTrans-retrieved AOT and FMF are generally more consistent with ground-based observations from AERONET than official AHI products, through three years of independent validation across the full-disk region. Specifically, AIRTrans achieves retrieval with RMSEs of 0.132 and 0.146 for AOT and FMF, respectively, compared to 0.216 and 0.284 for AHI products. AIRTrans shows a remarkable improvement on FMF, particularly in addressing the significant underestimation of the AHI products at over 90 % of individual sites. The performance of AIRTrans during two severe aerosol pollution events (intense dust storms and haze) further demonstrates its robust ability to capture spatiotemporal variations of AOT and FMF simultaneously.
