Ocean Trajectory Prediction Using Machine Learning Tools

Predicting ocean transport has many practical applications ranging from search and rescue operations to forecasting the spread of oil, marine debris, and biogeochemical tracers. Maritime environmental disasters like the 2010 Deepwater Horizon oil spill accentuate the importance of forecasting material transport in the ocean, yet the turbulent and chaotic nature of surface flow dynamics makes this an arduous task. This project considers an alternative to traditional theory-driven approaches to ocean forecasting by investigating data-driven methods for ocean transport prediction. The goal is to first learn from available data instead of prescribed laws of physics and then apply this information to new data. The first part explores whether simple artificial neural networks are capable of learning to predict 2-D particle trajectories using only previous velocity observations. The second part then evaluates how well these artificial neural networks and specialized social spatio-temporal graph convolutional neural networks are able to learn ocean trajectories from an unprecedented data set containing nearly 250 surface Lagrangian drifters released during the Grand Lagrangian Deployment in the Gulf of Mexico from July through October 2012. Collectively, this dissertation lays groundwork for developing strictly data-only machine learning forecasting methods for ocean material transport. Results herein demonstrate that while simple neural network architectures that have been around for decades may be able to learn trajectories generated by ocean general circulation models, they are unable to learn the large dynamical variability of real observed trajectories. At the same time, state-of-the-art neural networks that incorporate physical intuition show much greater potential at outperforming common regression models for forecasts on the order of days. These results show that problem-catered algorithms, together with sufficient observational data on which to train, may constitute valuable components within new advanced ocean prediction frameworks.