Sea level anomaly intelligent inversion model based on LSTM-RBF network

Abstract

Sea level anomaly (SLA) can be decomposed into two contributions: one from changes in mass in the water column (barotropic part) and the other from purely steric changes (baroclinic part) obtained by gravest empirical mode (GEM) lookup table. The establishment of GEM lookup table requires a large amount of data preprocessing (eliminating abnormal data, removing seasonal signals). The relationship between hydrological data (temperature, salinity and depth), vertical acoustic travel time and geopotential height is complex and requires human guidance. In view of the above shortcomings, this paper proposes the neural network-based SLA Intelligent Inversion model to fit the SLA model based on the traditional GEM field. The historical hydrological data is measured by conductivity temperature depth casts in the Eastern Taiwan and the northeastern part of Luzon. During the modeling process, firstly, the long short-term memory (LSTM) network is employed to detect and replace the abnormal data of the hydrological data. Secondly, the radial basis function (RBF) model based on the traditional GEM is constructed. The model can remove the seasonal signals. Besides, it can build the relationship between vertical acoustic travel time and geopotential height. The performance of the model is verified by the data set. It proves that compared with the traditional GEM, the relative error of network based on LSTM-RBF is less than 0.15%. Moreover, it can process abnormal data and remove seasonal signals. This study provides a novel method for preprocessing of hydrological data and SLA inversion based on Pressure-sensor-equipped inverted echo sounders.