Artificial Intelligence Technology-based Estimation Of Ocean Subsurface Temperature Structure

The ocean surface area accounts for 71%of the Earth's surface area,and its interior contains a large amount of information that humans have not yet discovered.Many important physical ocean phenomena and dynamical processes exist in a certain depth range below the sea surface.Therefore,it is of great significance to accurately obtain three-dimensional structure of the key parameters of the ocean subsurface,especially the thermo-thermal structure,for the study of the ocean circulation variability and climate change.However,the problem of insufficient measured data seriously restricts the knowledge of the ocean subsurface thermo-thermal structure.With the development of satellite observation technology,ocean science has ushered in the era of big data.With the help of the new generation of information technology,the combination of artificial intelligence technology and ocean theory to estimate the three-dimensional temperature field of the ocean subsurface has become one of the most advancing front and active research fields in current ocean science,which has important theoretical significance and application value.In this paper,based on multi-source ocean data,the key technology research of deep learning model in the estimation of ocean subsurface thermal structure is carried out.The main research work is as follows.(1)In this paper,a novel artificial neural network(ANN)model is proposed to estimate ocean subsurface temperature(OST)in the Western Pacific Ocean from multi-source data such as satellite remote sensing and Argo observations.Sea surface temperature(SST),sea surface height(SSH),sea surface salinity(SSS),the horizontal component(SSW_U)and vertical component(SSW_V)of sea surface wind(SSW)are selected as the input variables of the model.Based on the pre-processing of the data,the rasterized monthly average data from 2005-2015 and 2016 are selected as the training and test sets of the model,respectively.To verify the validity of the input parameters,five experimental schemes with different parameter combinations are designed in this paper.The experimental results show that the five sea surface parametershave a positive impact on the ANN estimation model,and the average root mean square error(RMSE)of the model with five input variables is 0.55°C.In addition,there is a significant signal of seasonal variation in the upper layer(above 200m)and the signal gradually decreases with the increase of depth.Furthermore,the estimation accuracy of the proposed model is better than that of existing estimation models such as random forest(RF),multiple linear regression(MLR)and e Xtreme Gradient Boosting(XGBoost).(2)For the problem of OST estimation in the global ocean,the ANN network structure is improved based on the network characteristics of Res Net and a new Gauss-Res NN hybrid estimation model combined with Gaussian mixture model(GMM)clustering algorithm is proposed to estimate the OST using the SST,SSS,SSH,SSW_U and SSW_V.The results show that the average RMSE andR~2 values of the model are 0.30°C and0.87,respectively.Compared with the existing hybrid estimation models,the clustering method chosen in this paper is more applicable to the classification of the global ocean,and the constructed Res NN structure is more effective than the ANN network in capturing the nonlinear motions in the ocean at medium and large scales.Moreover,the deeper the estimation depth is,the more obvious the advantage of the estimation model is.In addition,this paper also investigates the estimation performance of the model in different global ocean areas and different data sources,and verifies the effectiveness and robustness of the proposed model.(3)In order to solve the problem that OST estimation model is difficult to reveal the multi-scale nonlinear motion in the ocean,a new GR-XRAE hybrid estimation model with five input variables including the SST,SSS,SSH,SSWU and SSWV is designed to estimate the OST in the global ocean using the ideas of ensemble learning.The model is tested using global sea surface parameters from 2005-2015,and the effectiveness of the model is verified by Argo observation data.The results show that the average MSE andR~2 values of the model at all depths are 0.101°C and 0.899,respectively.Moreover,the estimation accuracy is better than the existing models such as the ANN model and SVM model,indicating that the model proposed in this paper can effectively solve the contradiction between the single nonlinear kernel function and the multi-scale nonlinear motion of the ocean.In addition,this paper also uses satellite remote sensing SST data instead of Argo observation data to test the GR-XRAE estimation model trained by Argo data.The results show that the GR-XEAE model has good generalizability.The research results of this paper can not only estimate the OST of various typical sea areas,but also provide technical support for the estimation of different oceanic elements such as the depth of the mixed layer,the depth of the barrier layer and the salinity of the subsurface layer.