A data-driven framework to estimate irrigation water withdrawal for hydrological modeling

Representing irrigation water withdrawal (IWW) in hydrological models is crucial for water management and climate change adaptation. However, the spatial and temporal variability of IWW remains poorly characterized. In this study, we present a data-driven framework to address this limitation. Using a dataset of annual IWW across river basins in France, we explicitly modeled the inter-catchment and the intra-catchment inter-annual variabilities of the annual IWW with random forest algorithms and a set of predictors, including monthly meteorological variables. Model robustness is evaluated through an extensive set of five cross-validation experiments. Prediction uncertainties are quantified using Generalized Additive Models for Location, Scale, and Shape (GAMLSS), combined with observed error structures. The model is then used to disaggregate annual IWW into monthly estimates using the SHapley Additive exPlanations (SHAP) values of monthly predictors. Our results show that, in a leave-one-out experiment, the IWW model achieves an r2 of 0.99. In contrast, removing temporal information strongly reduces the ability of the model to capture local inter-annual variability (r2 of 0.36). We also demonstrate the relevance of uncertainty estimation and identify the main determinants of irrigation in France, highlighting the contributions of the different predictors to the model’s predictions. The disaggregation to monthly IWW estimates is consistent with estimates from four process-based global models, as reported in the literature. Finally, this study delivers the CEMWat-Irrig-FR, a catchment-scale monthly IWW database in 936 catchments in France over the 2008-2022 period, along with a transferable methodology to estimate IWW for hydrological modeling.