A generative adversarial network–based unified model integrating bias correction and downscaling for global SST

This paper proposes a global sea surface temperature (SST) bias correction and downscaling unifying model based on the generative adversarial network. The generator of the model uses a bias correction module to correct the numerical model forecasting results. Then it uses a reusable shared downscaling module to improve the resolution of the corrected data gradually. The discriminator of the model evaluates the quality of the bias correction and downscaling results as a criterion for adversarial training. And a physics-informed dynamics penalty term is included in the adversarial loss function to improve the performance of the model. Based on the 1°-resolution SST forecasting results of the GFDL SPEAR (Seamless System for Prediction and Earth System Research) model, the authors select the Remote Sensing System observations as the refined targets and carry out validation experiments for three typical events at different scales and regions (ENSO, Indian Ocean dipole, and oceanic heatwave events). The model reduces the forecasting error by about 90.3% while increasing the resolution to 0.0625°×0.0625°, breaking the limitation of the resolution of the observation data, and the structural similarity with the observation results is as high as 96.46%.

摘要

本文提出了一种基于生成对抗网络的全球海表面温度(sea surface temperature, SST)偏差订正及降尺度整合模型. 该模型的生成器使用偏差订正模块将数值模式预测结果进行校正, 再用可复用的共享降尺度模块将订正后的数据分辨率逐次提高. 该模型的判别器可鉴别偏差订正及降尺度结果的质量, 以此为标准进行对抗训练. 同时, 在对抗损失函数中含有物理引导的动力学惩罚项以提高模型的性能. 本研究基于分辨率为1°的GFDL SPEAR模式的SST预测结果, 选择遥感系统(Remote Sensing System)的观测资料作为真值, 面向月尺度ENSO与IOD事件以及天尺度海洋热浪事件开展了验证试验: 模型在将分辨率提高到0.0625°×0.0625°的同时将预测误差减少约90.3%, 突破了观测数据分辨率的限制, 且与观测结果的结构相似性高达96.46%.