Inferring Ocean Transport Statistics With Probabilistic Neural Networks

Using a probabilistic neural network and Lagrangian observations from the Global Drifter Program, we model single particle transition probability density function (pdf) of ocean surface drifters. The pdf is represented by Gaussian mixture whose parameters (weights, means, covariances) are continuous functions latitude longitude determined to maximize likelihood observed drifter trajectories. This provides comprehensive description dynamics allowing for simulation trajectories estimation wealth dynamical statistics without need revisit raw data. As examples, compute global estimates mean displacements over 4 days lateral diffusivity. We use scoring rule compare our commonly used matrix models. Our outperforms others globally in three specific regions. A release experiment simulated using shows emergence concentrated clusters subtropical gyres, agreement with previous studies on formation garbage patches. An advantage that it continuous-in-space representation avoids discretize space, overcoming challenges dealing nonuniform approach, which embraces data-driven modeling, applicable many other problems fluid oceanography.