One Year Verification of Wave Predictions by Deep Learning Using Xception

Abstract

Accurate and timely wave forecast is vital for port maintenance and management as well as general maritime safety. Modern numerical wind-wave modeling calculates the physical processes involved in wave generation, interaction, propagation, and dissipation, requiring large amounts of computational power to timely complete the required calculations. Convolutional Neural Networks (CNNs) have shown to improve wave prediction in viewpoints of lesser computational cost and processing speed. This study examines the application of Xception deep learning architecture for wave predictions along Japanese coasts of the Sea of Japan with input features of the Japan Meteorological Agency's Grid Point Value Meso-Scale Model (GPV-MSM). In particular, this study aims to verify the forecast skill for a testing period of one year and to understand how wave forecasts by deep learning models can be improved and utilized.