深度学习降尺度算法对黄土高原地区极端降水的模拟能力评估

Performance assessment of deep learning downscaling algorithms in simulating extreme precipitation in the Chinese Loess Plateau region

关键词:  深度学习  统计降尺度  黄土高原  极端降水

Key words:  deep learning  statistical downscaling  Chinese Loess Plateau  extreme precipitation

摘要:

黄土高原 （CLP） 生态环境脆弱，作为气候敏感区，在全球变暖的背景下极端降水频率的提升将会增加该地区发生严重自然灾害的概率，对当地社会生产生活和人民生命财产安全造成威胁。目前，极端降水未来预估多采用全球气候模式 （GCM），其分辨率较低，对区域极端降水的模拟能力有限。鉴于此，文章将深度学习算法应用于黄土高原区域极端降水的降尺度，并对极端降水的模拟能力进行评估。 研究结果显示：卷积神经网络（CNN）、残差密集块网络（RDBNet）等深度学习模型能够很好地模拟黄土高原地区平均降水和极端降水强度。在极端降水的评价指标中，CNN1模型对R95p （日降水量＞95% 分位值的年累计降水量）、R99.5p（日降水量＞99.5%分位值的年累计降水量）和RX5day（年最大连续5 d 的降水量）的模拟能力最佳。研究结果客观揭示了深度学习模型在黄土高原地区极端降水模拟的适用性， 有助于进一步了解该地区未来降水变化情况并为极端降水的未来预估工作提供方法依据。同时，深度学习模型的应用也为未来研究气候变化提供了新的思路，提高对极端降水的模拟和预测能力。

Abstract:

Background, aim, and scope The Chinese Loess Plateau (CLP), recognized for its ecological vulnerability and sensitivity to climatic fluctuations, faces growing challenges under the influence of global warming. Notably, the rising occurrence of extreme precipitation events has emerged as a major threat to regional socio-economic development and public safety. While Global Climate Models (GCMs) are widely employed to forecast future precipitation trends, their limited spatial resolution often hampers their effectiveness in capturing localized hydrometeorological extremes. In response, this study explored the potential of advanced deep learning approaches to improve precipitation downscaling over the CLP, with a specific focus on evaluating their ability to replicate the characteristics of extreme rainfall events at regional scales. Materials and methods The research utilized high-resolution daily precipitation data derived from the IMERG (i.e., Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM)), featuring a temporal granularity of 0.5 h and a spatial resolution of 0.1°×0.1°. To inform the downscaling process, large-scale atmospheric predictors were sourced from the ERA5 reanalysis dataset, encompassing variables such as air temperature (T ), specific humidity (Q), geopotential height (Z ), meridional (U ) and, zonal (V ) wind velocities, provided at a coarser 2°×2° resolution. Precipitation downscaling was conducted through a comparative framework that integrated two deep learning architectures— Convolutional Neural Networks (CNN) and Residual Dense Block Networks (RDBNet)—alongside a traditional Generalized Linear Model (GLM). The dataset was partitioned into a calibration period (2001—2015) and a validation period (2016—2020) to assess model performance. Results Deep learning models exhibited a clear advantage in replicating both average precipitation patterns and the intensity of extreme rainfall events. Notably, the CNN1 architecture outperformed other models across multiple extreme precipitation indices, including R95p (annual total precipitation from days >95%), R99.5p (annual total precipitation from days >99.5%), and RX5day (maximum cumulative precipitation >5 days). Discussion The comparative analysis revealed distinct strengths and limitations among the modeling approaches. Deep learning techniques, particularly CNN and RDBNet, leveraged non-linear feature extraction capabilities that allow for more nuanced representation of complex precipitation dynamics. In contrast, the GLM, though interpretable, showed limited skill in replicating extreme event magnitudes. While all models demonstrated varying degrees of efficacy, the deep learning architectures consistently outshone the linear baseline, especially in simulating the intensity and temporal structure of highimpact precipitation episodes over the CLP. Conclusions The findings confirmed the superior performance of deep learning models in capturing extreme precipitation features across the CLP. By improving the spatial accuracy and sensitivity of downscaling processes, these approaches contributed to a more refined understanding of regional hydrological risks under climate change. The insights also established a methodological foundation for advancing precision in future climate projections and risk mitigation strategies. Recommendations and perspectives While this study highlighted the advantages of deep learning over linear models in the context of the CLP, the generalizability of these findings warrants further investigation. Future research should expand the scope to encompass diverse climatesensitive regions and explore a broader range of deep learning architectures. Furthermore, coupling deep learning downscaling with GCM-based future scenarios holds promise for enhancing regional precipitation forecasting under changing climate conditions.