Predicting sea surface salinity in a tidal estuary with machine learning

Type Article
Acceptance Date 2022-08 IN PRESS
Language English
Author(s) Guillou NicolasORCID1, Chapalain Georges1, Petton SébastienORCID2
Affiliation(s) 1 : Cerema, DTecREM, HA, Technopôle Brest-Iroise, Plouzané, France
2 : Ifremer, University of Brest, CNRS, IRD, LEMAR, Argenton, France
Source Oceanologia (0078-3234) (Elsevier BV) In Press
DOI 10.1016/j.oceano.2022.07.007
Keyword(s) Multilayer perceptron, Support vector regression, Random forest, River plume, Numerical model, Bay of Brest

As an indicator of exchanges between watersheds, rivers and coastal seas, salinity may provide valuable information about the exposure, ecological health and robustness of marine ecosystems, including especially estuaries. The temporal variations of salinity are traditionally approached with numerical models based on a physical description of hydrodynamic and hydrological processes. However, as these models require large computational resources, such an approach is, in practice, rarely considered for rapid turnaround predictions as requested by engineering and operational applications dealing with the ecological monitoring of estuaries. As an alternative efficient and rapid solution, we investigated here the potential of machine learning algorithms to mimic the non-linear complex relationships between salinity and a series of input parameters (such as tide-induced free-surface elevation, river discharges and wind velocity). Beyond regression methods, the attention was dedicated to popular machine learning approaches including MultiLayer Perceptron, Support Vector Regression and Random Forest. These algorithms were applied to six-year observations of sea surface salinity at the mouth of the Elorn estuary (bay of Brest, western Brittany, France) and compared to predictions from an advanced ecological numerical model. In spite of simple input data, machine learning algorithms reproduced the seasonal and semi-diurnal variations of sea surface salinity characterised by noticeable tide-induced modulations and low-salinity events during the winter period. Support Vector Regression provided the best estimations of surface salinity, improving especially predictions from the advanced numerical model during low-salinity events. This promotes the exploitation of machine learning algorithms as a complementary tool to process-based physical models.

Full Text
File Pages Size Access
Publisher's official version IN PRESS 15 2 MB Open access
Top of the page