A multi-task deep learning framework for enhancing cloud-top height retrieval accuracy

doi:10.1080/17538947.2025.2605411

A multi-task deep learning framework for enhancing cloud-top height retrieval accuracy

2026-01-02

0

PRE

AI

DOI：10.1080/17538947.2025.2605411

Original

Original request for help

Save

Abstract

En 中文

Cloud-top height (CTH) is a fundamental parameter that influences the radiative effects of clouds and plays a critical role in improving the prediction accuracy of numerical weather prediction models. This study introduces a deep learning framework, namely MultiTask-CNN, for the simultaneous retrieval of CTH, cloud phase and multilayer information based on satellite remote sensing data. The datasets used in this study span from January to June 2018. Leveraging multi-source data from observation of the Advanced Geostationary Radiation Imager(AGRI) onboard FY-4A, ERA5 reanalysis, and the combined lidar-radar cloud profiles (joint product of CALIOP and CPR), the model learned complex interdependencies among cloud properties to enhance the retrieval accuracy of CTH. Compared with traditional models (CNN, XGBoost and TwoStage-CNN), the Multi-Task-CNN improved accuracy. Specifically, the RMSE decreased from 2.08 km (CNN), 1.88 km (TwoStage-CNN) and 1.75 km (XGBoost) to 1.70 km, MAE was reduced to 0.79 km, and R-2 slightly increased to 0.90. This improvement was particularly pronounced for multi-layer and high-level clouds. Ablation studies further highlight the benefits of incorporating cloud phase and multilayer information in enhancing model robustness. This study underscores the potential of multi-task learning in cloud property retrieval, offering valuable insights for improving climate and weather prediction.

Keywords:

Cloud top height

deep learning

multi-task learning

cloud phase

cloud multilayer information

A multi-task deep learning framework for enhancing cloud-top height retrieval accuracy

Abstract

Journal

Organization