Remote Sensing Inversion Models Of Sea Surface Salinity Based On Machine Learning Method

Remote sensing technology is an effective tool to obtain accurate parameters of water quality and perform dynamic simulation and analysis at large scales. The salinity is an important ocean physicochemical parameter. It is of great importance to study the salinity distribution and variation, which is helpful for the understanding of the ocean characteristics and function in the ocean-atmosphere system. In this study, we elaborated the mechanism and methods of sea surface salinity inversion, and built sea surface salinity inversion model based on machine learning. Furthermore, we simulated the sea surface salinity at large-scale dynamically.In this study, we selected the typical Hong Kong Estuary and its adjacent sea as the study area, the satellite image data was Landsat 8, which collected from January to December, 2014. The field data set included all sorts of water quality parameters was collected from 2000 to 2014. Firstly, we investigated and analyzed the sensitive parameters of the sea surface salinity deeply and then established the sea surface salinity inversion model based on Deep Learning, Kernel Ridge Regression and the Support Vector Regression. The main research work and conclusion is as follows:(1) We selected the total nitrogen, the total phosphorus and the temperature which had high correlation with the sea surface salinity as the sensitive factors, this conclusion was drew based on the analysis of the Person correlation between the sea surface salinity and the chlorophyll a, suspended solids, total nitrogen, total phosphorus and the temperature by the measured water salinity parameters of Hong Kong waters on 2000-2014.(2) Establishing the high-precision inverse model of the sensitive sea salinity water quality parameters suited for the Hong Kong waters. As for the total nitrogen, the highest accuracy was achieved in the exponential form based on the sensitive combination bands B1-B3, with 2R =0.948. While for the total phosphorus, the third multinomial regression equations fitting better: where 2R =0.945. Meanwhile, the sea surface temperature inverse model was established based on the radiative transfer equation.(3)Establishing the sea surface salinity inversion model based on Deep Learning(DNN), Kernel Ridge Regression(KRR) and the Support Vector Regression(SVR). The retrieval accuracy was considerable through training and optimizing different model. The best result was achieved in the optimized DNN sea surface model by testing different models: R =0.845, MSE =2.38, MAE =1.126;The KRR model had the highest precision based on the rbf kernel function where g =0.1, a =0.464 with a fitting precision as follows: R =0.851, MSE =2.93, MAE =1.135;While the best predict result of the SVR model based on the rbf model was achieved where the C=215, g =0.017, and the fitting precision are: R =0.840, MSE =3.043, MAE =1.126.By studying and analyzing different machine learning method, we had found that it is feasible to establish the sea surface salinity using the sensitive factors as the input, and the result was satisfying and reliable. No matter the deep neural network(DNN) model or the regression model based on the kernel function such as KRR and SVR obtained high-precision except the obvious difference in time efficiency.