On Friday, November 3rd, we will have an OCP Seminar given by Dr. Julien Le Sommer, a CNRS senior research scientist in the field of Computational Oceanography.
This will be a hybrid in-person/zoom seminar taking place in the Monell Auditorium. Please email the event contact for the zoom information. The title and abstract are provided down below.
Title: Probabilistic predictions of Lagrangian drift at the ocean surface using machine learning
Abstract: Predicting the drift of objects lost at sea is key to several operational applications such as marine pollution, oil spills, and floating wreckage. It is also key better understanding the transport pathways which govern the exchange of water masses and biogenic matter in the upper ocean. In practice, existing operational systems are based on tracking schemes, which leverage estimates of ocean surface currents and other environmental data (winds, waves) for reconstructing the drift of lagrangian particles. But, because of the chaotic nature of lagrangian dynamics, small uncertainties in the surface current, in the initial position or in the drift function can lead to large error in the predicted drift. This is why many operational systems use a combination of stochastic models and ensemble methods for accounting for uncertainty in lagrangian drift reconstructions. The lack of direct measurements of ocean surface currents in the open ocean is also problematic for practical applications, and it is not clear how to leverage ancillary data, as for instance sea surface temperature or ocean color data for lagrangian applications. In this talk, I will present results from a recently developed deep learning framework that directly models the evolution of the probability distribution of presence of objects lost at sea from various ocean surface data. I will illustrate the performance of this framework with Observing System Simulation Experiments in both coastal and open ocean environments, and show how it allow to perform retrospective reconstructions of lagrangian trajectories with uncertainties from indirect and corrupted ocean data.
