Climate-driven changes in Venice Lagoon hydrography under global warming scenarios

Coastal lagoons are highly dynamic transitional systems whose hydrographic properties are strongly modulated by atmospheric forcing, freshwater inputs, and exchanges with the open sea, reflecting coupled land–sea–atmosphere processes across coastal interfaces. Accurately simulating their temperature and salinity variability remains challenging, particularly under climate change scenarios, due to the high computational cost of process-based hydrodynamic models and the limited availability of long observational time series. Here, we present a data-driven modelling framework to reproduce and project monthly surface water temperature and salinity in the Venice Lagoon, one of the most complex and vulnerable coastal systems in the Mediterranean region, using a Convolutional Neural Network (CNN). The model is trained using irregular monthly observations collected between 2001 and 2004 at three representative stations (marine, intermediate, and riverine), combined with a minimal set of physically interpretable atmospheric and oceanographic predictors, including 2 m air temperature, precipitation, mean sea level, and offshore sea surface salinity. Despite the short training period, the CNN accurately reproduces the observed seasonal and interannual variability, achieving high skill scores (R² > 0.96 for temperature and R² > 0.85 for salinity at most stations). A sensitivity analysis reveals distinct dominant drivers across the lagoon, with oceanic forcing prevailing near the inlets and atmospheric–terrestrial controls becoming increasingly important in river-influenced areas. The validated model is subsequently employed to explore synthetic climate change scenarios corresponding to 1.5, 2, and 3 °C global warming levels relative to pre-industrial conditions. Results indicate a pronounced amplification of the seasonal cycle, with summer surface water temperature increases exceeding 6 °C and salinity increases above 4 PSU at the riverine station under the 3 °C scenario. These changes suggest substantial future alterations of lagoon hydrography, with potential implications for ecosystem functioning and resilience. Overall, this study demonstrates the potential of CNN-based approaches as computationally efficient tools for climate impact assessment in complex coastal environments, complementing traditional hydrodynamic models and enabling rapid scenario exploration.