Assessing climate change impacts on wetlands inundation: a comparison of multiple regression and artificial neural networks hybrid models

Wetlands are highly vulnerable to climate-driven hydrological changes, particularly in semi-arid Mediterranean regions. This study evaluated the impact of climate change on wetland inundated area by comparing two hybrid modelling approaches: a multiple linear regression approach (MULREG) and a nonlinear autoregressive neural network with exogenous inputs (NARX). Both integrate aquifer discharge data from a physics-based groundwater flow model (MODFLOW). The framework combines Landsat-derived inundation images (1984–2015), produced via supervised near-infrared classification, hydroclimatological drivers, and derived variables, including potential evapotranspiration, effective precipitation, and aquifer discharge. Although some NARX realizations achieved lower prediction errors than MULREG, the average performance across all realizations was comparable or slightly poorer, suggesting no clear model superiority. Future scenarios were generated using a delta-change method driven by an ensemble of ISIMIP GCM projections under multiple RCPs, sampling uncertainty via + 1 °C and + 3 °C warming levels, and contrasting regional precipitation changes. Scenario propagation revealed significant differences between the models: MULREG produced comparatively muted reductions in wetland inundated area, even under severe hydroclimatic shifts, whereas NARX yielded a stronger response to projected changes in the hydroclimatological drivers and produced results that were more consistent with the sensitivities observed during the historical period. Under the most extreme scenario (+ 3 °C with reduced precipitation), the mean inundated area decreased by ~ 28% (from 4.31 km² to 3.09 km²), with larger reductions under dry conditions. The results highlight the need for appropriate model selection when translating climate scenarios into wetland responses and support the use of nonlinear memory-aware models for groundwater-dependent systems.