Estimation Of Ocean Subsurface Temperature And Salinity Structure In The South China Sea Based On Artificial Intelligence Technology

The South China Sea,as the largest marginal sea in the northwest Pacific Ocean,contains many important physical ocean phenomena and dynamical processes.Accurate acquisition of information on the subsurface temperature and salt structure of the South China Sea plays an important role in understanding the ecosystem and climate change.With the development of satellite remote sensing technology and artificial intelligence technology,it is of great theoretical significance and application value to carry out inversion studies of subsurface thermohaline structure in the South China Sea using artificial intelligence methods based on multi-source ocean data.In this paper,we used multi-source sea surface parameters such as sea surface temperature(SST),sea surface salinity(SSS),sea surface height(SSH),sea surface wind(SSW)observed by remote sensing satellites and Argo data.Based on Deep Forest algorithm and light gradient boosting Machine(Light GBM)algorithm,an estimated model with neural network ensemble was proposed to estimate the ocean subsurface thermal structure(OSTS)in the South China Sea and an improved Light GBM-based Deep Forest(LGB＿DF)model was proposed to estimate the ocean subsurface salinity structure(OSSS)in the South China Sea.The root mean square error(RMSE),normalized root mean square error(NRMSE)and coefficient of determination(R~2)were used to verify the reliability of the model estimates.The main research contents and results are as follows.(1)Study on the estimation of OSTS in the South China Sea based on ensemble learning ideas combined with neural networksThe subsurface thermal structure of the South China Sea was inferred from the multi-source satellite remote sensing data and Argo measurements from January 2010to December 2020,and an ensemble model was built by integrating the Deep Forest algorithm and Light GBM algorithm with neural networks after data pre-processing.The results show that the proposed estimation model performs well in estimating the OSTS in the South China Sea and the estimation accuracy is better than that of the single model and the Support Vector Regression(SVR)model.The regional average RMSE and R~2values were 0.0509psu and 0.8962,respectively.In terms of spatial distribution,the OSTS estimated by the estimation model has good spatial agreement and coincidence with the horizontal spatial distribution of Argo measured at equal depths of 50,100,500and 1000 m.The model-estimated OSTS of the South China Sea has a consistent vertical distribution with Argo measurements when viewed on the southern continental slope of China,in the West Luzon vortex region,in the central southern part,in the East Vietnam vortex region,and on a cross section that slopes across the South China Sea from southwest to northeast.(2)Study on the estimation of OSSS in the South China Sea based on LGB＿DF modelThe multi-source satellite remote sensing data and Argo measurements from January 2010 to December 2020 were selected.An improved Deep Forest model based on the Light GBM method(LGB＿DF)model was developed for the estimation of the OSSS in the South China Sea after data pre-processing.The results show that the LGB＿DF model is reliable and has a better estimation accuracy than the Deep Forest model and the SVR model.The regional average RMSE and R~2 values were 0.3325℃and 0.8880,respectively.The horizontal spatial distribution indicates that the estimated salinity based on the LGB＿DF model and the Argo measured salinity show a consistent spatial distribution in 2020.The performance metrics indicate that the performance of the LGB＿DF model varies with depth,tending to show better performance in the shallow layers of the ocean,probably because the physical state is more easily described relative to the sea surface.The model performance also varies with season,with the salinity model having the highest estimation accuracy in autumn(NRMSE=0.0593,R~2=0.9259).Lower NRMSE and higher R~2values in all seasons indicate a good seasonal applicability.The scatter distribution indicates that most of the scatter points are uniformly and densely distributed on the contours.In addition,a comparison experiment with different combinations of sea surface parameters was set up to compare the estimation of OSSS in the South China Sea,which proved that sea surface parameters and geographic information are necessary parameters for accurate estimation of the OSSS in the South China Sea.