Xiao et al. (2025) High-resolution ensemble retrieval of cloud properties for all-day based on geostationary satellite

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2025
• Date: 2025-12-12
• Authors: Haixia Xiao, Feng Zhang, Lingxiao Wang, Baoxiang Pan, Yannian Zhu, Minghuai Wang, Wenwen Li, Bin Guo, Jun Li
• DOI: 10.1038/s41612-025-01263-x

Research Groups

• Key Laboratory of Polar Atmosphere-Ocean-Ice System for Weather and Climate of Ministry of Education/ Shanghai Key Laboratory of Ocean-Land-Atmosphere Boundary Dynamics and Climate Change, Department of Atmospheric and Oceanic Sciences & Institutes of Atmospheric Sciences, Fudan University, Shanghai, China
• Nanjing Innovation Institute for Atmospheric Sciences, Chinese Academy of Meteorological Sciences–Jiangsu Meteorological Service, Nanjing, China
• Jiangsu Key Laboratory of Severe Storm Disaster Risk/Key Laboratory of Transportation Meteorology of CMA, Nanjing, China
• RIKEN Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS), RIKEN, Wako, Saitama, Japan
• Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China
• School of Atmospheric Sciences, Nanjing University, Nanjing, China
• Engineering Research Center of Optical Instrument and System, the Ministry of Education, Shanghai Key Laboratory of Modern Optical System, University of Shanghai for Science and Technology, Shanghai, China
• National Satellite Meteorological Center, China Meteorological Administration, Beijing, China

Short Summary

This study introduces CloudDiff, a novel generative diffusion model, for high-resolution (1 km) and all-day ensemble retrieval of cloud properties (Cloud Optical Thickness, Cloud Effective Radius, Cloud Top Height, Cloud Phase) from geostationary satellite data, providing uncertainty quantification and significantly improving retrieval accuracy and reliability compared to deterministic methods.

Objective

• To develop a novel cloud property retrieval method based on a generative diffusion model (CloudDiff) that can generate high spatiotemporal resolution (1 km, 10 min) cloud properties for both daytime and nighttime conditions, while also quantifying retrieval uncertainty.

Study Configuration

• Spatial Scale: Cloud properties retrieved at 1 km resolution from 2 km thermal infrared observations.
• Temporal Scale: All-day retrieval, with Himawari-8 AHI capturing full-disk images every 10 minutes. Training data collected from 2016 to 2018.

Methodology and Data

• Models used:
  • CloudDiff: Conditional diffusion model with a UNet architecture incorporating multi-head attention modules.
  • Deterministic Model: Supervised learning approach using a UNet architecture (for comparison).
• Data sources:
  • Input: Himawari-8 Advanced Himawari Imager (AHI) 2 km thermal infrared (TIR) radiance measurements (8 bands: 6.95 µm, 7.35 µm, 8.60 µm, 9.63 µm, 10.45 µm, 11.20 µm, 12.35 µm, and 13.30 µm) and viewing zenith angle (VZA).
  • Target/Ground Truth: Moderate Resolution Imaging Spectroradiometer (MODIS) Level-2 cloud products (MOD06 L2 and MYD06 L2) for Cloud Optical Thickness (COT), Cloud Effective Radius (CER), Cloud Top Height (CTH), and Cloud Phase (CLP) at 1 km spatial resolution.
  • Validation/Comparison: Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) Level-2 cloud layer products (CTH and CLP) at 1 km horizontal resolution, and Himawari-8 operational cloud products (CTH and CLP).

Main Results

• CloudDiff successfully retrieves high-resolution (1 km) cloud properties (COT, CER, CTH) and identifies cloud phase (CLP) for all-day conditions.
• Ensemble averaging of 30 samples significantly improves retrieval accuracy and reliability compared to single samples and a deterministic UNet model.
  • For CLP identification, CloudDiff ensemble mean achieved an Overall Accuracy (OA) of 84.55%, outperforming the deterministic model (OA 83.25%).
  • For COT, the ensemble mean (30 samples) yielded a Mean Absolute Error (MAE) of 6.12 and a Root Mean Squared Error (RMSE) of 12.43, compared to the deterministic model's MAE of 6.85 and RMSE of 13.28.
  • For CER, the ensemble mean (30 samples) achieved an MAE of 5.70 µm and an RMSE of 8.83 µm, compared to the deterministic model's MAE of 6.32 µm and RMSE of 9.51 µm.
  • For CTH, the ensemble mean (30 samples) achieved an MAE of 1.29 km and an RMSE of 2.32 km against MODIS. When validated against CALIPSO, CloudDiff CTH showed highest agreement with MAE of 1.73 km and RMSE of 2.81 km, slightly outperforming the deterministic model and Himawari-8 operational products.
• CloudDiff demonstrates strong structural similarity (SSIM = 0.68) and good pixel-level accuracy (PSNR = 25.87) for COT retrievals at an ensemble size of 30.
• Single samples from CloudDiff provide sharper images and better capture local details than the ensemble mean, while the ensemble mean still outperforms deterministic retrievals in representing cloud structures.
• Uncertainty quantification (ensemble spread) reflects retrieval-related errors but does not fully account for intrinsic biases inherited from MODIS products.
• Application to Typhoon In-Fa demonstrated CloudDiff's ability to provide sharper, more localized 1 km cloud properties and valuable uncertainty estimates during extreme weather events, with minimal unphysical negative values (COT 1.57%, CER 0.19%, CTH 0.18%).

Contributions

• Introduces the first cloud property retrieval model using geostationary satellites to achieve high-resolution (1 km) ensemble retrievals across all-day conditions.
• Presents a novel generative diffusion model (CloudDiff) for cloud remote sensing, enabling the quantification of retrieval uncertainties, a critical advancement over existing deterministic methods.
• Provides high spatiotemporal resolution (1 km, 10 min) and all-day cloud properties (COT, CER, CTH, CLP) with improved accuracy and reliability through ensemble averaging.
• Offers a versatile methodology that can be adapted for the retrieval of other atmospheric variables from satellite observations, promoting the application of generative artificial intelligence in atmospheric science.

Funding

• National Natural Science Foundation of China (42222506, 42450254, 12147101)
• JSPS KAKENHI Grant No. 25H01560
• JST-BOOST Grant No. JPMJBY24H9

Citation

@article{Xiao2025Highresolution,
  author = {Xiao, Haixia and Zhang, Feng and Wang, Lingxiao and Pan, Baoxiang and Zhu, Yannian and Wang, Minghuai and Li, Wenwen and Guo, Bin and Li, Jun},
  title = {High-resolution ensemble retrieval of cloud properties for all-day based on geostationary satellite},
  journal = {npj Climate and Atmospheric Science},
  year = {2025},
  doi = {10.1038/s41612-025-01263-x},
  url = {https://doi.org/10.1038/s41612-025-01263-x}
}