Retrieving Ocean Subsurface Temperature Anomaly Based On Multisource Satellite Observations

The ocean as the most important part of the Earth has played an essential role in the global climate variability and change by heat uptake and storage.The current satellite remote sensing can obtain multi-scale sea surface observations,but it cannot detect the key dynamic information inside the ocean.Many subsurface phenomena have surface manifestations that can be interpreted with the help of satellite measurements.Therefore,the key dynamic parameters in the ocean interior can be retrieved from satellite observations of sea surface parameters by setting up appropriate remote sensing inversion models.This paper proposed novel approaches based on machine learning(SVM,RF)and geographically weighted regression(GWR)models to estimate subsurface temperature anomaly(STA)over large-scale ocean from satellite measurements including sea surface height anomaly(SSHA),sea surface temperature anomaly(SSTA),sea surface salinity anomaly(SSSA),and sea surface wind anomaly(SSWA)with the help of Argo in situ data for model training and testing.The main research contents and results are as follows:1)We proposed a support vector machine(SVM)method to estimate subsurface temperature anomaly(STA)in the global ocean.The results were reliable with reasonable accuracy by validating using the Argo STA data.The average R2and MSE of the 15levels are 0.485/0.457/0.443 and 0.0087/0.0086/0.0090 for 4-attributes(SSHA,SSTA,SSSA,SSWA)/3-attributes(SSHA,SSTA,SSSA)/2-attributes(SSHA,SSTA)SVR,respectively,suggesting SSS and SSW in addition to SSH and SST are positive input parameters for SVR prediction.SVR model can well retrieve STA in the global ocean.2)We also proposed a novel method for retrieving STA in the global ocean based on random forest(RF).The results indicated the estimated accuracy in the January(Winter),2010 was the best,and the average R2 and MSE are0.661 and 0.0073 respectively,followed by July(Summer),October(Fall)and April(Spring),and the overall accuracy still keep high.Moreover,the model performed best in the Indian Ocean with highest accuracy(R2and MSE are 0.778 and 0.0416 respectively).RF method was suitable for estimating subsurface temperature anomaly in different seasons of 2010 over the global ocean.3)We evaluated SVR and RF model from three different aspects in regard to different parameter input combinations,different season applications and estimation error distribution.The results showed RF outperformed SVR in global STA estimation in all those aspects,and RF model was more suitable for the global STA detection than SVR model.4)We developed a new satellite-based geographically weighted regression(GWR)model to estimate the STA in the Indian Ocean.Due to the poor interpretation of the machine learning model,we used the GWR model with local regression coefficient visualization to extract the STA.The results demonstrated GWR model can accurately retrieve STA in the Indian Ocean by combined use of surface observations and Argo in situ data.The average R2 and MSE are 0.781 and 0.0119 respectively for the model performance measure.In addition,compared to the other three input factors,the SSHA which played a leading role on the model had the most significant influence on the GWR weight determination.The mean of the regression coefficients for the four input factors about SSHA,SSTA,SSWA,and SSSA are 0.617,0.089,0.077,and 0.041 respectively.Compared to the global ordinary least square(OLS)model,the GWR model had an advantage in the local modeling,and could retrieve STA more accurately on large scale by considering the spatial nonstationarity feature.