A data-driven framework for estimating monthly irrigation water withdrawals at the catchment scale

Exploring future hydrological conditions and supporting decision-makers require datasets of irrigation water withdrawals (IWW) that capture spatial, interannual, and seasonal variabilities. Most databases do not meet these criteria, as they are typically limited to specific regions and time periods and are often available only at the annual time scale. The quantification of IWW has recently received increased attention through machine-learning approaches that can model the complex relationship between IWW and explanatory factors. However, methodological challenges remain in temporally disaggregating annual IWW data and in assessing the robustness and uncertainty of machine-learning-based estimates.

In this work, we present a comprehensive data-driven framework for estimating monthly IWW at the catchment scale. Our approach allows for the interpolation and extrapolation of IWW, for uncertainty quantification, and for the temporal disaggregation of annual IWW to a monthly resolution. Interpolation is performed using Random Forest (RF) algorithms, which are evaluated using five spatio-temporal cross-validation experiments. Prediction uncertainty distributions are modeled using Generalized Additive Models for Location Scale and Shape (GAMLSS) and observed error structures. Extrapolation of annual IWW is then achieved using a Generalized Additive Model (GAM). Finally, annual IWW data are disaggregated to monthly values using the contribution of monthly meteorological and soil wetness predictors.

The methodology is applied to 656 French catchments using the French Water Withdrawals National Database (BNPE), which makes available annual IWW values since 2008 at the local administrative level, and open-source predictors, such as area equipped for irrigation, crop type and monthly meteorological data. RF models achieve high predictive skill (r² ≈ 0.99), but performance declines sharply under spatio-temporal data removal (r² ≈ 0.4), underscoring the the importance of rigorous validation and comprehensive uncertainty quantification. A SHapley Additive exPlanations (SHAP) analysis reveals physically consistent relationships between predictors and their contribution to model predictions. Monthly disaggregated IWW are consistent with seasonal patterns simulated by four global hydrological models, with peak values occurring in summer. We demonstrate the value of the modelling framework at the catchment scale by extending the dataset backward in time and by projecting IWW into the future.

This work received funding from the European Life Revers'Eau project.