UNET-vs-Analogues_data

In Europe, temperature variations are mainly driven by the North Atlantic atmospheric circulation. Here we investigate a convolutional neural network (a UNET) for reconstructing daily temperature anomalies in Europe from sea level pressure as a proxy of the atmospheric circulation, and we compare the results with the traditional analogues approach. We show an excellent ability of the UNET to estimate temperature variations given information from sea level pressure only. This novel method outperforms the analogue method, at both daily and inter-annual time scales. Our study also shows that during the training, the UNET learns information such as the seasonal cycle of the relationship between sea-level pressure and temperature anomalies, which could explain part of its excellent scores. This work opens up promising prospects for estimating the contribution of atmospheric variability to temperature variations.

在欧洲区域，气温变化主要受北大西洋大气环流驱动。本研究针对利用海平面气压作为大气环流的替代指标，重建欧洲区域逐日气温距平的任务，探究了卷积神经网络（UNET）的应用效果，并将其结果与传统相似方法进行了对比。实验结果表明，仅依靠海平面气压输入信息，UNET便可出色地估算气温变化。该新型方法在逐日及年际时间尺度上均优于传统相似方法。本研究同时发现，在训练过程中，UNET能够学习到海平面气压与气温距平之间关系的季节循环特征，这一特性可部分解释其优异的模型表现。本研究为估算大气变率对气温变化的贡献提供了极具前景的研究方向。