2026-07-17 合肥物質科学研究院（HFIPS）
Global AOD distribution maps. Global AOD distribution in (a) February, (b) May, (c) August, and (d) November. (Image by HUANG Honglian)
＜関連情報＞
- https://english.hf.cas.cn/nr/rn/202607/t20260717_1178563.html
- https://ieeexplore.ieee.org/document/11577750
注意機構強化型コルモゴロフ・アーノルドネットワークを用いたGaofen-5B POSPからのエアロゾル光学的厚さの推定 Aerosol Optical Depth Retrieval From Gaofen-5B POSP Using Attention-Enhanced Kolmogorov–Arnold Networks
Zhuochi Liu; Honglian Huang; Xiao Liu; Rufang Ti; Zhijun Chen; Gengke Wang;…
IEEE Transactions on Geoscience and Remote Sensing Published:24 June 2026
DOI:https://doi.org/10.1109/TGRS.2026.3706811
Abstract:
Atmospheric aerosols play a crucial role in governing Earth’s radiation balance and significantly impact air quality, climate change, and human health. The complexity of their chemical composition and the heterogeneity of their spatial distribution complicate the accurate retrieval of aerosol optical depth (AOD). In recent years, deep-learning-based AOD retrieval methods have shown great potential, but effectively exploring spectral, angular, and polarimetric information remains challenging. In this research, we propose a novel AOD retrieval model, the attention Kolmogorov–Arnold network (AKAN), which integrates residual channel attention modules with Kolmogorov–Arnold networks (KANs). The algorithm utilizes data from the Particulate Observing Scanning Polarimeter (POSP) onboard Chinese Gaofen-5B (GF-5B) satellite. By combining one year of POSP observations with ground-based measurements from global Aerosol Robotic Network (AERONET) sites and supplemental regional sun-photometer observations, we constructed a dataset of 243 584 matched samples that characterize the complex nonlinear relationships between polarimetric signatures and AOD. AKAN demonstrates excellent performance in AOD retrieval, achieving high accuracy with R2=0.9336 , root-mean-square error ( RMSE=0.0492 ), mean absolute error ( MAE=0.0262 ), and 94.25% of retrievals falling within the expected error (EE) envelope. Notably, SHapley Additive exPlanations (SHAP) were employed to examine model interpretability and the consistency between feature contributions and retrieval behavior. The results indicate that AKAN not only captures the statistical relationships between spectral-polarimetric information and AOD but also provides physically meaningful insights into multiband polarimetric characteristics. Furthermore, cross comparisons with Moderate Resolution Imaging Spectroradiometer (MODIS) AOD products validate the superior performance of the proposed model. Case studies of pollution events and analyses of global aerosol distributions demonstrate the model’s strong spatial–temporal monitoring capability and robust generalizability under diverse environmental conditions.
