AUTHOR=Xiang Xiaohua , Li Yongxuan , Wu Xiaoling , Liu Zhu , Wu Lei , Wu Biqiong , Jin Chuanxin , Zeng Zhiqiang 

TITLE=Future variation and uncertainty source decomposition in deep learning bias-corrected CMIP6 global extreme precipitation historical simulation

JOURNAL=Frontiers in Earth Science

VOLUME=Volume 13 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2025.1601615

DOI=10.3389/feart.2025.1601615

ISSN=2296-6463

ABSTRACT=Global circulation models (GCMs) serve as pivotal tools in climate science research. Despite their critical role in understanding and predicting climate change, GCMs often exhibit significant discrepancies with observational data due to systematic and random errors, which has driven the progress of bias correction (BC) techniques. This study explores a bias correction approach based on convolutional neural networks (CNNs) to improve the accuracy of Expert Team on Climate Change Detection and Indices (ETCCDI) extreme precipitation indices calculated from the Coupled Model Intercomparison Project Phase Six (CMIP6) daily predictions. Specifically, this research employs historical period data (1950–2014) for eight ETCCDI extreme precipitation indices from 10 GCMs to train eight individual CNN-based bias correction models, using the HadEX3 reference dataset for evaluation. All corrected data showing mean absolute percentage error (MAPE) were consistently reduced to below 0.1. Subsequently, these well-trained models are further utilized to predict ETCCDI extreme precipitation for the future under four Shared Socioeconomic Pathway (SSP) scenarios, and the projections of extreme precipitation changes are investigated across global continents. In addition, this study endeavors to separate and quantify three different components of uncertainty (model uncertainty, scenario uncertainty, and internal variability) associated with ETCCDI extreme precipitation indices and evaluate the impact of bias correction on uncertainty variation. The results indicate that CNNs are effective in correcting historical precipitation extremes. In the future period, extreme precipitation shows an increasing trend in general. The degree of change in R10mm is relatively small and reaches its peak in the medium term, whereas the variation in Rx1day is more pronounced and increases over time. Further analysis reveals that model uncertainty is the predominant source of uncertainty in ETCCDI extreme precipitation indices, accounting for more than 80% of total uncertainty. Implementation of CNNs as a BC method could significantly reduce model uncertainty but at the cost of increasing the proportion of scenario uncertainty and internal variability. This research not only highlights the potential of the CNN-based deep learning technique in enhancing the accuracy and reliability of extreme precipitation predictions but also provides insights into uncertainty decomposition and variation to better understand various sources of uncertainty within climate projections.