Deep Learning Prediction Methods For Sea Surface Temperature Based On Spatiotemporal Multidimensional Influences

In global climate researches,marine ecosystem researches,and ocean-related applications,it is of considerable significance to accurately and effectively observe and predict Sea Surface Temperature(SST).However,various physical and environmental factors affect the changes in SST,making it highly random and uncertain.Therefore,it is still a challenge to propose a highly accurate SST prediction method.SST prediction methods based on the temporal influence often focus on capturing the static influence of historical SST but ignore the dynamic influence in the prediction process,so these methods meet the performance bottlenecks.In order to capture the static and dynamic influence effectively and enhance the ability to process the temporal information,this paper proposed the Gated Recurrent Unit(GRU)Encoder-Decoder(GRU Encoder–Decoder,GED)to predict SST in the Bohai Sea and the South China Sea.Besides,in order to fuse the multidimensional spatiotemporal influence and further improve the accuracy of the SST prediction,this paper also proposed the Convolutional GRU with Multilayer Perceptron(CGMP)to predict SST in the Bohai Sea and the South China Sea.The main research contents and innovations of this paper are as follows:(1)The GED with SST codes and Dynamic Influence Link(DIL)was proposed,which considered both the static and dynamic influence.On the one hand,the SST code is calculated by all the hidden states of the GRU encoder based on the attention mechanism.It is also associated with each predicted future SST.So that it can capture the static influence more effectively.On the other hand,DIL links SST codes,the early predicted future SST,and the unpredicted future SST.So that it can capture the dynamic influence more effectively.In this paper,experiments were conducted at three different schemas(daily mean,weekly mean,and monthly mean)and various prediction scales in the Bohai Sea and the South China Sea.Results showed that the prediction performance of GED was better than that of other classic models in different sea areas,different schemas,and different prediction scales.Especially in the long-scale and monthly mean SST prediction,the prediction performance advantage of GED was the most obvious.Besides,the optimal observation window size that balanced the prediction and time performance,along with the attention relationship between historical SST and future SST were obtained.(2)The spatiotemporal CGMP was proposed,which considered the neighbor influence.The convolutional layer of CGMP can capture the neighbor influence more effectively in the spatial dimension,making up for the shortcomings of the models that are based on the temporal influence and do not consider the spatial influence.The GRU layer and multilayer perceptron layer of CGMP can process historical information more effectively in the temporal dimension.In this paper,experiments were also conducted at three different schemas(daily mean,weekly mean,and monthly mean)and various prediction scales in the Bohai Sea and the South China Sea.Results showed that the prediction performance of CGMP was better than that of other comparison models in different sea areas,different schemas,and different prediction scales.Its prediction performance has been further improved compared to GED in the weekly and monthly mean SST prediction.Besides,the law of error distribution in the Bohai Sea daily mean SST prediction was explored.