FOWD: A Free Ocean Wave Dataset for Data Mining and Machine Learning

TL;DR

This paper introduces FOWD, a comprehensive, cleaned ocean wave dataset and processing framework designed to support data mining and machine learning research on rogue waves, including a Python toolkit and analysis of predictive features.

Contribution

The paper presents FOWD, a new large-scale ocean wave dataset with a processing framework and toolkit, enabling advanced analysis of rogue wave phenomena using machine learning.

Findings

Surface elevation kurtosis and maximum wave height are strong predictors for rogue waves.

Crest-trough correlation, spectral bandwidth, and mean period are key predictors when only spectral data is available.

Processed over 4 billion waves from the CDIP buoy data catalogue.

Abstract

The occurrence of extreme (rogue) waves in the ocean is for the most part still shrouded in mystery, as the rare nature of these events makes them difficult to analyze with traditional methods. Modern data mining and machine learning methods provide a promising way out, but they typically rely on the availability of massive amounts of well-cleaned data. To facilitate the application of such data-hungry methods to surface ocean waves, we developed FOWD, a freely available wave dataset and processing framework. FOWD describes the conversion of raw observations into a catalogue that maps characteristic sea state parameters to observed wave quantities. Specifically, we employ a running window approach that respects the non-stationary nature of the oceans, and extensive quality control to reduce bias in the resulting dataset. We also supply a reference Python implementation of the FOWD processing toolkit, which we use to process the entire CDIP buoy data catalogue containing over 4 billion waves. In a first experiment, we find that, when the full elevation time series is available, surface elevation kurtosis and maximum wave height are the strongest univariate predictors for rogue wave activity. When just a spectrum is given, crest-trough correlation, spectral bandwidth, and mean period fill this role.