Ensemble Approach to Deep Learning and Numerical Ocean Modelling of Adriatic Storm Surges

Storm surges are among the most serious threats to Venice, Chioggia, Piran and other historic coastal towns in Northern Adriatic. Adriatic Sea has a well defined lowest seiche period of approximately 22 hours and its amplitude decays on the scale of several days, reinforcing (or diminishing) the tidal signal, depending on the relative phase lag between tides and surges. This makes prediction of Adriatic sea level extremely difficult using conventional deterministic models. The current state-of-the-art predictions of sea surface height (SSH) hence involve numerical ocean models using ensemble forcing. These simulations are computationally-demanding and time consuming, making the method unsuitable for operational or civil rescue services with limited access to dedicated high-performance computing facilities.

Ensemble approach to deep learning offers a possible solution to the challenges described above. Even though training a deep network may involve substantial computational resources, the subsequent forecasting -- even ensemble forecasting -- is fast and delivers near-realtime SSH predictions (and associated error variances) on a personal computer. In this work we present an ensemble SSH forecast using new deep convolutional neural network for sea-level prediction in the Adriatic basin and compare it to the standard approach using state-of-the-art publicly available modelling components (NEMO ocean circulation model and TensorFlow libraries for deep learning).