基于ConvGRU和自注意力的海表温度偏差订正研究

海表温度（Sea Surface Temperature， SST） 是一个重要的海洋物理量， 其准确预报对于水产养殖和预测海洋环境信息等海洋相关领域的研究至关重要。数值预报是目前预报海表温度的一种常用方法， 但数值预报模型所产生的预报结果往往与实际观测数据有偏差， 因此有必要对数值预报产品的偏差进行订正。本文提出了一种结合ConvGRU 神经网络与自注意力（Self-Attention， SA） 的新型时空混合海表温度订正模型（ConvGRU-SA）， 对南海海表温度预报数据进行偏差订正， 该模型适用于利用卫星遥感数据对海表温度数值预报产品进行订正。经与ConvLSTM、ConvGRU 等网络模型对比， 证明了ConvGRU-SA 模型的优越性， 设置不同的超参数进行实验， 提高模型订正准确率。订正后该区域的海表温度预报均方根误差从0.52 ℃降低至0.32 ℃， 准确率提高了38.4%， 优于现有模型。

Study of SST Bias Revision Based on ConvGRU and Self-Attention

Sea surface temperature (SST) is an important physical quantity of the ocean, and accurate forecasting of SST is crucial for research in marine-related fields such as aquaculture and predicting information about the marine environment, and numerical prediction is now a common method for predicting SST. However, the prediction results produced by numerical prediction models often deviate from the actual observations, so it is necessary to revise the deviation of numerical prediction products. In this paper, a new spatio-temporal hybrid SST revision model (ConvGRU-SA) is proposed to be constructed by combining ConvGRU neural network and the attention mechanism to revise the deviation of SST forecast data in the South China Sea, which is suitable for revising the numerical SST prediction products using satellite remote sensing data. Comparison with network models such as ConvLSTM and ConvGRU proves the superiority of the ConvGRUSA model, and different hyper-parameters are set to conduct experiments to improve the model revision accuracy. The root-mean-square error (RMSE) of the region is reduced from 0.52 ℃ to 0.32 ℃ after the revision, and the accuracy is improved by 38.4%.