The robustness of conceptual rainfall-runoff modelling under climate variability-A review

Conceptual rainfall-runoff (CRR) models are widely used tools in climate change impact studies. However, the assumption that hydroclimate variables are stationary is no longer justifiable when input forcing is significantly different from the hydro-climatological conditions used in model building. It is particularly important to identify and discard such modelling that are unsuitable for future prediction in a calibration/evaluation strategy. Previous literatures have thoroughly investigated the implications of climate change on catchments around the world, but a few studies have systematically assessed the transferability of CRR models. In this paper, the transferability of CRR models in a climate variability context is reviewed. First, the development of the data split methods for examining parameter dependence on climate and the associated objective function with model robustness metrics are presented. Second, by comparing the outcomes collectively, both the robustness assessment of the classic differential split-sample test (DSST) and its variants, such as the large-sample generalized split-sample test (GSST), and linkages between model transferability with non-stationary climate and with catchment characteristics are explored. Among others, we answer the following questions: (1) Under which climatic constraints can models empirically be transferred? (2) Are models more difficult to transfer in catchments with certain characteristics? A set of model transferability criteria that explicitly consider potential failure scenarios at different steps in a data splitting approach is established. Thus a strategy to diagnose model transferability is proposed for routinely assessing the prediction ability of CRR models under various climate conditions, particularly when results are used to inform adaptation decision-making.