Reconstruction Of The Three-dimensional Temperature Field Of The Upper Ocean Considering The Characteristics Of The Fron

The oceanic front is a mesoscale environmental feature with complex temperature structures.In-depth investigation of its temperature structures can better utilize oceanic resources and promote marine economic development.However,due to the inadequate number of available temperature profiles,reconstructing the three-dimensional temperature field based on sea surface observations has been proved to be reliable in most regions worldwide.This paper examines the reconstruction method of the upper ocean three-dimensional temperature field considering the characteristics of the oceanic front based on remote sensing data and field information.The main contents and conclusions are as follows:(1)Four reconstruction models were built using linear and polynomial regression with sea surface temperature(SST)and sea level anomaly(SLA)inputs.The polynomial model with SST and SLA had the best performance.An analysis of the relationship between model errors and SST gradient showed that areas of decreased accuracy are spatially similar to areas of high SST gradient.Model errors exhibit quasi-linear growth trends relative to SST gradient,with the growth rate changing with depth.The impact of the front on accuracy increases and then decreases with depth,peaking near 250 m.Therefore,considering the impact of the front on vertical temperature profile reconstruction is necessary to improve reconstruction accuracy.(2)A neural network model was employed to investigate the improvement of temperature profile reconstruction performance by introducing SST gradient as input.Model sensitivity evaluation indicated that normalized inputs and the Adam optimizer were beneficial for model training.Five sets of comparative models were designed to compare model accuracy with and without SST gradient inputs,and the results showed that the model error was smaller with SST gradient inputs and significantly improved below 200 m.Additionally,the use of nonlinear activation functions and deeper model structures also improved accuracy.(3)A reconstruction model considering regional SST gradients was constructed by combining convolutional neural networks and neural networks and incorporating the influence of the environmental temperature field.Sensitivity testing showed that the model achieved the highest accuracy with a learning rate and decay factor of 1e-05.The model had high accuracy at all depths,with correlation coefficients above 0.95.Compared with the neural network model,the RMSE of the model was reduced by 2.62% at 50 m and 7.21% at 250 m,and the error was also reduced at other depths,demonstrating the importance of considering regional features.The reconstructed temperature field had a smaller deviation compared to ARMOR3 D,further proving the validity of the model.