Spatio-Temporal Deep Learning for Ocean Current Prediction Based on HF Radar Data

Abstract

Ocean surface current prediction is necessary to carry a variety of marine activities, such as disaster monitoring, search and rescue operations, etc. There are three traditional forecasting approaches: (i) numerical based approach, (ii) time series based approach and (iii) machine learning based approach. Unfortunately, their prediction accuracy was limited since they did not cooperate with spatial and temporal effects together. In this paper, we present a novel current prediction model, which is a combination between Convolutional Neural Network (CNN) to extract spatial characteristic and Gated Recurrent Unit (GRU) to find a relationship of temporal characteristic. The dataset is collected by high frequency (HF) radar station's located along coastal Thailand's gulf by GISTDA from 2014 to 2016. It was an intensive experiment comparing our method and eight existing methods, e.g., ARIMA, kNN, Perceptron, Multilayer Perceptron (MLP), etc. The results show that our network outperforms almost all baselines in terms of RMSE for 11.21% and 27.01% averaging improvement on U and V components, consecutively.