Long-term numerical modeling of sea level in the Northern Adriatic for a machine learning downscaling system

Abstract

<p>Urban settlements near to coastal environments are exposed to ocean and cryosphere change, such as sea level rise and extreme sea levels. High-resolution sea level prediction systems have become fundamental tools for taking preventive measures in the face of extreme events, mainly in the most vulnerable coastal locations. Techniques such as Machine Learning (ML) are at the forefront of the development in this sector, as they can reduce the computational time needed to reproduce the results of costly high resolution dynamic models. In this line, different authors have reported results for the prediction of oceanographic variables using ML approaches (Camus et al., 2019; Costa et al., 2020; Zust et al., 2021), mainly for significant wave height, sea level and surge component of sea level. Generally, these works use global and/or regional databases as training data for ML tools.</p> <p>With the aim of developing a data-driven system for sea level downscaling, by means of very high-resolution circulation model output used as a training data for a ML framework, in this work the results of a long-term numerical modeling of sea level are presented, carried out in the Northern Adriatic. The numerical model implemented correspond to SURF-SHYFEM, a 3-D finite element hydrodynamic model that solves the primitive equations under hydrostatic and Boussinesq approximations. As atmospheric forcing, mean sea level pressure, and meridional and zonal components of wind speed have been included, both from the ERA5 database. For the boundary conditions, sea level has been considered from two databases, the Copernicus Mediterranean Forecasting System (available from November 2020 to present, with tides included in sea level) and the Copernicus Mediterranean Sea Physics Reanalysis (available from 1987 to June 2021, without tides in sea level). Both databases were used on initial analysis in the representation of surge component of sea level when tides are or not included in the boundary condition. The validation of the results has been carried out by comparison with tide gauges located near the Venice Lagoon, from ISPRA[1] and PSMSL[2].</p> <p>The results show that the model reproduces accurately the sea level (correlation 94% and RMSE 0.09 [m]) and the surge component of sea level (correlation 91% and RMSE 0.06 [m]) measured at the location of the tide gauge. The next step will consist of using such output as a training set for ML-based techniques, with the aim of developing an accurate and cost-effective downscaling tool.</p> <div><br /> <div> <p>[1] Istituto Superiore per la Protezione e la Ricerca Ambientale. Available at: https://www.mareografico.it/</p> </div> <div> <p>[2] Permanent Service for Mean Sea Level. Available at: https://psmsl.org/</p> <p> </p> <p><strong> REFERENCES</strong></p> <p>Camus, P., Herrera, S., Guitiérrez, J.M. and Losada, I.J. (2019). Statistical downscaling of seasonal wave forecast. Ocean Modelling 138, 1-12.</p> <p>Costa, W., Idier, D., Rohmer, J., Menendez, M. and Camus, P. (2020). Statistical prediction of extreme storm surges based on a fully supervised weather-type downscaling model. J. Mar. Sci. Eng. 8, 1028.</p> <p>Zust, L., Fettich, A., Kristan, M. and Licer, M. (2021). HIDRA 1.0: Deep-Learning-Based ensemble sea level forecasting in the Northern Adriatic. Geosci. Model. Dev. 14, 2057-2074.</p> </div> </div>