Application of the neural network on the GNSS-Reflectometry data for the estimation of the significant wave height

Abstract

Globally, remote sensing is the best way to estimate Significant Wave Height (SWH). Traditional sensors such as satellite altimeters are generally used to provide SWH. However, due to poor temporal resolution and poor signal quality during heavy rain, altimeters signals are not suitable during heavy rain condition. To overcome above limitations, Global Navigation Satellite System-Reflectometry (GNSS-R) is widely used to generate the ocean parameters. However, due to the poor range resolution of GNSS-R signals, GNSS-R requires complex algorithm to generate SWH. In this paper, a Neural Network (NN) based machine learning technique is proposed to estimate the SWH using Cyclone GNSS (CYGNSS) observables. Levenberg-Marquardt algorithm is applied to train and update the weight function and bias of the network. Optimum number of layers and nodes in each layer are selected by the criteria which minimize the error of the output of the NN. Once the network is formed, the estimated SWH is validated using SWH of WW-3 model. The training data gives the Correlation Coefficients (CC) equals to 0.91 and Root Mean Square Difference (RMSD) equals to 0.35 m. The validation data gives the RMSD and the Mean Error (ME) equals to 0.36 m and -0.002 m respectively. NN output is also compared with Jason-3 altimeter SWH data. The analysis shows that similar to altimeter, GNSS-R signals can also be used to generate SWH.