Research Of Ocean Temperature Prediction Based On Argo Data

Ocean temperature is one of the most important factors in hydrology research.The master of marine temperature's space-time distribution and variation is of great importance when it comes to the research of meteorology,navigation,acoustic.The complexity and uncertainty of marine environment,which make the prediction of temperature difficult.In order to better predict ocean temperature,this paper presents a predicting method both spatially and temporally.Given such demands,this paper proposes a ocean temperature predicting method based on SVR with high spatial resolution.Spatially,the distribution of ocean temperature has characteristics of wide ranges and large scales.The key to the prediction is how to find the fitting function and use this function to do specific prediction.Argo data is adopted and SVR ocean temperature model is constructed.When constructing the SVR training model,this paper use cross validation and the grid search method to optimize the model parameters.Based on the optimal SVR model,the original data sets with the resolution of 1°×1° are then improved to a finer granularity,which resolution is 0.1°×0.1°.The model is visualized in this paper.Compared with linear regression and k-neighbor regression,the SVR model based on RBF kernel function is better.Temporally,a method consists of time domain feature sequence and LSTM neural network is proposed.The method analyzes temporal characteristics of ocean temperature time series.Self-correlation,seasonal fluctuation and tendency of the time series are included.And the time series are reconstructed.LSTM model is constructed according to the reconstructed time sequence.The experimental results revealed that ocean temperature time series reconstructing method based on time domain features can not only reduce the length of the previous sequence which is used to predict,but to some extent improve the predictive ability of this LSTM model.The method shows satisfactory predicting performance regardless of the difficulties arising from the deepness of water deepen.Methods above have been testified by the Argo data and turn out to be feasible both spatially and temporally.The proposed spatio-temporal predicting method is suitable for other waters.