Xiao et al. (2025) Reconstructing historical cloud ice water fraction using machine learning and multi-source satellite data from 1983 to 2009

Identification

• Journal: Atmospheric Research
• Year: 2025
• Date: 2025-10-17
• Authors: Dezhi Xiao, Chunsong Lu, Sinan Gao, Jiashan Zhu, Yang Li, Ru Zhou, Junjun Li, Jing Yang, Naifu Shao
• DOI: 10.1016/j.atmosres.2025.108571

Research Groups

• China Meteorological Administration Aerosol-Cloud and Precipitation Key Laboratory & Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters (CIC-FEMD), Nanjing University of Information Science and Technology, Nanjing, China
• Guangzhou Institute of Tropical and Marine Meteorology of China Meteorological Administration, GBA Academy of Meteorological Research, Guangzhou, China
• Guangdong Meteorological Data Centre, Guangzhou, 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
• Beijing Weather Modification Center, Beijing, China
• National Institute of Education (NIE), Nanyang Technological University (NTU), Singapore

Short Summary

This study develops and evaluates machine learning schemes to reconstruct global monthly mean ice water fraction (IWF) from 1983 to 2009, leveraging multi-source satellite and reanalysis data to correct biases in earlier observations and provide an improved historical dataset for cloud-climate interaction studies.

Objective

• To obtain more accurate historical global monthly mean cloud ice water fraction (IWF) data for the period 1983-2009 by developing and applying machine learning schemes.
• To evaluate the predictive accuracy, computational efficiency, and interpretability of various machine learning algorithms for IWF prediction.

Study Configuration

• Spatial Scale: Global
• Temporal Scale: Reconstruction from July 1983 to December 2009.

Methodology and Data

• Models used: Light Gradient Boosting Machine (LGB), eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Multilayer Perceptron (MLP). SHapley Additive exPlanations (SHAP) for interpretability. A new metric, R2-Time score (RTS), was introduced for comprehensive evaluation.
• Data sources: CloudSat-CALIPSO satellite observations, International Satellite Cloud Climatology Project (ISCCP) dataset, fifth-generation Earth Reanalysis Atmospheric (ERA5) reanalysis data.

Main Results

• All four machine learning schemes demonstrated high predictive accuracy on the test set for global monthly mean IWF.
• The Light Gradient Boosting Machine (LGB) scheme exhibited superior computational efficiency and was selected for the final long-term reconstruction and SHAP analysis.
• SHAP analysis indicated that specific humidity, latitude, and relative humidity are the primary factors influencing IWF prediction, and the LGB scheme's prediction mechanism aligns with physical expectations without strong temporal or spatial dependency.
• The reconstructed global monthly mean IWF dataset for July 1983 to December 2009 successfully captures distinct annual cycles and reproduces climatic variability consistent with ISCCP and CloudSat-CALIPSO records.
• The reconstructed dataset effectively corrects systematic biases in cloud-phase classification present in the original ISCCP data.

Contributions

• Development and validation of robust machine learning methodologies for reconstructing historical global monthly mean ice water fraction, addressing limitations in early satellite observations.
• Introduction of the R2-Time score (RTS) as a novel metric for comprehensively evaluating machine learning model performance in terms of both accuracy and computational efficiency.
• Provision of an improved, long-term (1983-2009) historical global monthly mean IWF dataset, offering enhanced data quality for investigations into cloud-climate interactions.
• Demonstration of the physical consistency and interpretability of the selected machine learning model (LGB) using SHAP analysis.

Funding

Not explicitly stated in the provided text.

Citation

@article{Xiao2025Reconstructing,
  author = {Xiao, Dezhi and Lu, Chunsong and Gao, Sinan and Zhu, Jiashan and Li, Yang and Zhou, Ru and Li, Junjun and Yang, Jing and Shao, Naifu},
  title = {Reconstructing historical cloud ice water fraction using machine learning and multi-source satellite data from 1983 to 2009},
  journal = {Atmospheric Research},
  year = {2025},
  doi = {10.1016/j.atmosres.2025.108571},
  url = {https://doi.org/10.1016/j.atmosres.2025.108571}
}