A deep learning–based U-Net model for ENSO-related precipitation responses to sea surface temperature anomalies over the tropical Pacific

Abstract

SST–precipitation feedback plays an important role in ENSO evolution over the tropical Pacific and thus it is critically important to realistically represent precipitation-induced feedback for accurate simulations and predictions of ENSO. Typically, in hybrid coupled modeling for ENSO predictions, statistical atmospheric models are adopted to determine linear precipitation responses to interannual SST anomalies. However, in current coupled climate models, the observed precipitation–SST relationship is not well represented. In this study, a data-driven deep learning-based U-Net model was used to construct a nonlinear response model of interannual precipitation variability to SST anomalies. It was found that the U-Net model outperformed the traditional EOF-based method in calculating the precipitation variability. Particularly over the western-central tropical Pacific, the mean-square error (MSE) of the precipitation estimates in the U-Net model was smaller than that in the EOF model. The performance of the U-Net model was further improved when additional tendency information on SST and precipitation variability was also introduced as input variables, leading to a pronounced MSE reduction over the ITCZ.

摘要

SST–降水反馈过程在热带太平洋ENSO演变过程中起着重要作用, 能否真实地在数值模式中表征SST–降水年际异常之间的关系及相关反馈过程, 对于准确模拟和预测ENSO至关重要. 例如, 在一些模拟ENSO的混合型耦合模式中, 通常采用大气统计模型 (如经验正交函数; EOF) 来表征降水 (海气界面淡水通量的一个重要分量) 对SST年际异常的线性响应. 然而在当前的耦合模式中, 真实观测到的降水–SST统计关系还不能被很好地再现出来, 从而引起 ENSO模拟误差和不确定性. 在本研究中, 使用基于深度学习的U-Net模型来构建热带太平洋降水异常场对SST年际异常的非线性响应模型. 研究发现: U-Net模型的性能优于传统的基于EOF方法的模型. 特别是在热带西太平洋海区, U-Net模型估算的降水误差远小于EOF模型的模拟. 此外, 当SST和降水异常的趋势信息作为输入变量也被同时引入以进一步约束模式训练时, U-Net模型的性能可以进一步提高, 如能使热带辐合带区域的误差显著降低.