In the last decades, advancements in computational science have greatly expanded the use of artificial neural networks (ANNs) in hydrogeology, including applications on groundwater forecast, variable selection, extended lead-times, and regime-specific analysis. However, ANN-model performance often omits the sensitivity to observational uncertainties in hydroclimate forcings. The goal of this paper is to implement a data-driven modeling framework for assessing the sensitivity of ANN-based groundwater forecasts to the uncertainties in observational inputs across space, time, and hydrological regimes. The objectives are two-folded. The first objective is to couple an ANN model with the PAWN sensitivity analysis (SA). The second objective is to evaluate the scale- and process-dependent sensitivities of groundwater forecasts to hydroclimate inputs, computing the sensitivity index in groundwater wells (1) across the whole time-series (for the global sensitivity analysis); (2) across the output sub-regions with conditions of water deficit and water surplus (for the ‘regional’ sensitivity analysis); and (3) at each time step (for the time-varying sensitivity analysis). The implementation of the ANN-PAWN occurs in 68 wells across the Northern High Plains aquifer, USA, with pre-time-step rainfall, evapotranspiration, snowmelt, streamflow, and groundwater measurements as inputs. Results show that evapotranspiration and rainfall are the major sources of uncertainty, with the latter being particularly relevant in water surplus conditions and the former in water deficit conditions. The time-varying sensitivity analysis leads to the identification of localized sensitivities to other sources of uncertainty, as snowmelt in spring or river flow during the annual peak period at the groundwater level.

The aim of this paper is to improve semiseasonal forecast of groundwater availability in response to climate variables, surface water availability, groundwater level variations, and human water management using a two-step data-driven modeling approach. First, we implement an ensemble of artificial neural networks (ANNs) for the 300 wells across the High Plains aquifer (USA). The modeling framework includes a method to choose the most relevant input variables and time lags; an assessment of the effect of exogenous variables on the predictive capabilities of models; and the estimation of the forecast skill based on the Nash-Sutcliffe efficiency (NSE) index, the normalized root mean square error, and the coefficient of determination (R²). Then, for the ANNs with low- accuracy, a MultiModel Combination (MuMoC) based on a hybrid of ANN and an instance-based learning method is applied. MuMoC uses forecasts from neighboring wells to improve the accuracy of ANNs. An exhaustive-search optimization algorithm is employed to select the best neighboring wells based on the cross correlation and predictive accuracy criteria. The results show high average ANN forecasting skills across the aquifer (average NSE > 0.9). Spatially distributed metrics of performance showed also higher error in areas of strong interaction between hydrometeorological forcings, irrigation intensity, and the aquifer. In those areas, the integration of the spatial information into MuMoC leads to an improvement of the model accuracy (NSE increased by 0.12), with peaks higher than 0.3 when the optimization objectives for selecting the neighbors were maximized.tT.