Seamless transition of altimetric retracked sea levels using neural network technique: Case study using simulated data

Abstract

Waveform retracking has become a standard data processing protocol to optimize the estimation of sea level, particularly over coastal oceans. In the proximity of land, combining different retracking algorithms are essential for dealing with high diversity of altimetric waveform patterns. However, retrackers cannot be simply switched to another due to the existence of offset among retrackers. The existence of offset value creates 'a jump' in the sea level profiles, thus reducing the precision of the estimated sea level parameter. In this paper, neural network technique is explored to reduce the offset values, and to produce a seamless transition of sea level when switching retrackers. The analysis is conducted over 100,000 simulated data based on Monte Carlo simulation. The experiment includes six sets of varying parameters (i.e. number of hidden layer, algorithms in hidden and output layers, and training algorithm). The results indicate that the neural network (set 2) with six hidden layers, algorithms of Logsig and Tansig for hidden and output layers, respectively, and Levenberg-Marquardt for training algorithm is the best parameters for offset reduction. It has the highest correlation, and the lowest root mean square error and standard deviation of difference, giving it the best rank when compared to the other five sets.