U-Net: A deep-learning method for improving summer precipitation forecasts in China

Abstract

A deep-learning method named U-Net was applied to improve the skill in forecasting summer (June–August) precipitation for at a one-month lead during the period 1981–2020 in China. The variables of geopotential height, soil moisture, sea level pressure, sea surface temperature, ocean salinity, and snow were considered as the model input to revise the seasonal prediction of the Climate Forecast System, version 2 (CFSv2). Results showed that on average U-Net reduced the root-mean-square error of the original CFSv2 prediction by 49.7% and 42.7% for the validation and testing set, respectively. The most improved areas were Northwest, Southwest, and Southeast China. The anomaly same sign percentages and temporal and spatial correlation coefficients did not present significant improvement but maintained the comparable performances of CFSv2. Sensitivity experiments showed that soil moisture is the most crucial factor in predicting summer rainfall in China, followed by geopotential height. Due to its advantages in handling small training dataset sizes, U-Net is a promising deep-learning method for seasonal rainfall prediction.

摘要

本研究应用了名为U-Net的深度学习方法来提高中国夏季 (6–8月) 降水的预报技能, 预报时段为1981–2020年, 预报提前期为一个月. 将位势高度场, 土壤湿度, 海平面气压, 海表面温度, 海洋盐度和青藏高原积雪等变量作为模型输入, 本文对美国NCAR气候预报系统第2版 (CFSv2) 的季节性预报结果进行了修正. 结果显示, 在验证集和测试集上, U-Net平均将原CFSv2预测的均方根误差分别减少了49.7%和42.7%. 预报结果改善最大的地区是中国的西北,西南和东南地区. 然而, 同号率和时空相关系数没有得到明显改善, 但仍与CFSv2的预测技巧持平. 敏感性实验表明, 土壤湿度是预测中国夏季降雨的最关键因素, 其次是位势高度场. 本研究显示了U-Net模型在训练小样本数据集方面的优势, 为我国汛期季节性降雨预测提供了一种有效的深度学习方法.