Autoregressive denoising diffusion for predicting trajectories of floating objects in oceans

Motivation The increasing amount of plastic debris in the oceans calls for quick action to prevent irreversibly damaging our world’s largest ecosystem. To this end, tracking plastic debris and understanding its dynamics could facilitate collection campaigns and help monitor the evolution of the threat. To achieve this goal, accurate models are necessary to predict the dynamics of floating objects at the ocean surface, which are subject to currents and winds. Physical models and remote sensing data estimate these influencing forces. However, using them directly in process-based models still leads to a significant gap between the true dynamics and the predicted trajectory. Hence, we aim to minimize this gap by resorting to data-driven machine-learning methods.

Data We can identify two different scenarios where the dynamics of floating objects differ: trajectories close to coastal regions and trajectories in the open ocean. As a consequence, we focus on two different datasets: the first aims to predict dynamics in coastal regions for 24 hours. The second focuses on open-ocean dynamics, where we try to predict trajectories for multiple days. As target variables, we use data from the Global Drifter program, which contains several thousand GPS-tracked free-floating buoys. The contextual information about the ocean surface current is extracted from Copernicus Marine and HYCOM. Wind data is taken from ERA5.

Approach We develop a denoising diffusion model that generates multiple trajectories based on surface current and wind, as provided by physical models. In contrast to the unstructured i.i.d. Gaussian noise in standard denoising diffusion, we use a more suitable process: Brownian motion noise, which has a small variance close to the start of the trajectories and increases with time. The denoiser model is autoregressive and based on a multilayer recurrent neural network that iteratively learns to remove the noise from random realizations of this Brownian motion.

Results We found that the model not only outperforms physical models on the coastal dataset but also provides a posterior distribution of the predicted trajectories, thus offering a measure of uncertainty without additional overhead.