Which spatial discretization for which distributed hydrological model?

Abstract. Distributed hydrological models are valuable tools to derive distributed estimation of water balance components or to study the impact of land-use or climate change on water resources and water quality. In these models, the choice of an appropriate spatial scale for the modelling units is a crucial issue. It is obviously linked to the available data and their scale, but not only. For a given catchment and a given data set, the "optimal" spatial discretization should be different according to the problem to be solved and the objectives of the modelling. Thus a flexible methodology is needed, especially for large catchments, to derive modelling units by performing suitable trade-off between available data, the dominant hydrological processes, their representation scale and the modelling objectives. In order to represent catchment heterogeneity efficiently according to the modelling goals, and the availability of the input data, we propose to use nested discretization, starting from a hierarchy of sub-catchments, linked by the river network topology. If consistent with the modelling objectives, the active hydrological processes and data availability, sub-catchment variability can be described using a finer nested discretization. The latter takes into account different geophysical factors such as topography, land-use, pedology, but also suitable hydrological discontinuities such as ditches, hedges, dams, etc. For small catchments, the landscape features such as agricultural fields, buildings, hedges, river reaches can be represented explicitly, as well as the water pathways between them. For larger catchments, such a representation is not feasible and simplification is necessary. For the sub-catchments discretization in these large catchments, we propose a flexible methodology based on the principles of landscape classification, using reference zones. These principles are independent from the catchment size. They allow to keep suitable features which are required in the catchment description in order to fulfil a specific modelling objective. The method leads to unstructured and homogeneous areas within the sub-catchments, which can be used as modelling units. It avoids map smoothing by suppressing the smallest units, the role of which can be very important in hydrology, and provides a confidence map (the distance map) for the classification. The confidence map can be used for further uncertainty analysis of modelling results. The final discretization remains consistent with the scale of input data and that of the source maps. We present an illustration of the method using available data from the upper Saône catchment (11 700 km2) in France. We compare the results with more traditional mapping approach, according to the landscape representation and input data scale.