Neural network modelling of temperature and salinity in the Venice Lagoon

This study applies an artificial neural network (ANN) to simulate monthly temperature and salinity variations at three stations in the Venice lagoon, which have been selected to represent different regimes (marine, riverine and intermediate) in terms of relevance of local processes and exchanges with the open sea. Four key predictors are shown to play a major role: mean offshore sea level, 2 m air temperature, precipitation for the lagoon water temperature, integrated with offshore sea surface salinity for the lagoon water salinity. The development of the ANN is based on only 4 years of observations, taken irregularly over time with an approximately monthly frequency. Despite this, the ANN achieves an accurate reproduction of both variables with large R2 and reasonably small, normalized root-mean-square errors at all stations, except for the salinity at the marine station, where the model presents a spurious variability, which is absent in observations. Sensitivity analysis shows that the 2 m air temperature is the dominant predictor for water temperature while sea-level and sea surface salinity are the principal predictor of salinity fluctuations, with precipitation exerting a relevant role mainly at the riverine station. The ANN has been used for a set of synthetic climate change analyses considering 1.5, 2 and 3 degrees C global warming levels with respect to preindustrial levels. An overall warming of lagoon water with maximum increase in summer is expected (up to 6 degrees C in the 3 degrees C global warming level), resulting in an amplification of the annual cycle amplitude. The expected increases in salinity have a strong gradient across the lagoon, are largest at the riverine station, and (analogously to the changes in temperature) amplify the salinity annual cycle amplitude.