A data-driven framework for the temporal extrapolation of annual water withdrawals for hydrological modelling

In the context of anthropogenic climate change and increasing pressure on water resources from human use, it is necessary to provide stakeholders with tools to quantify water availability under present and future conditions, and to guide public policy in water management. To this end, anthropogenic effects need to be integrated into hydrological modeling. One of the major challenges in modeling human-impacted hydrological systems is the quantification of water withdrawals at the appropriate temporal and spatial scales. Due to a general lack of direct observational data, these withdrawals must often be modeled. The strategy for data-based modeling of water withdrawals depends on the water use sector: irrigation is traditionally subject to a process-based approach, while public freshwater supply is often modelled using regression techniques. Recently, machine-learning techniques have been explored to model freshwater withdrawals and, in the irrigation sector, to identify drivers and, in rarer cases, to predict water withdrawals.   

In this study, we present a data-driven framework to quantify irrigation water with limited data.  We illustrate our methodological development with an application over 74 non-nested catchments in France, where water withdrawals are documented based on declarations for a short historic period (since 2008) and at a coarse temporal resolution (annual volumes). To obtain longer time series for the calibration of a hydrological model, we perform a temporal extrapolation of irrigation water withdrawals at the catchment scale. To predict the annual withdrawal, we use a mixed-effects model that explicitly distinguishes between structural variation (e.g. annual change in area equipped for irrigation) and random variation (e.g. change in meteorological and soil conditions). These two terms are modeled using a random forest algorithm. We evaluate the robustness of the model by excluding, at a turn, from the training set: (i) catchments located in the same region to evaluate the spatial extrapolation performance, and (ii) a year of data for all the catchments to evaluate the temporal extrapolation. Our results show that the structural variation modelling term is particularly robust on temporal extrapolation (overall RMSE of 25% of the predicted value), while the random variation modelling term performs well in both temporal and spatial extrapolation (overall Pearson correlation coefficient of 0.72 and 0.80). We discuss how the framework can be used to disaggregate annual values of water withdrawal and be integrated into hydrological modelling.

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