Towards a methodology for testing models as hypotheses in the inexact sciences

Abstract

This paper reviews the issues involved in treating hydrology as an example of an inexact science faced with significant epistemic uncertainties. It proposes a novel method for developing limits of acceptability for testing hydrological models as hypotheses about how a catchment hydrological system might function. The approach is based only on an analysis of the available observations and the consideration of event mass balance for successive rainfall-runoff events. It is shown that there are many events that are subject to epistemic uncertainties in the input data so that mass balance is not satisfied. The proposed approach allows taking these epistemic uncertainties into account in a pragmatic way before any model runs are made. It is an approach that might be applicable in other areas of environmental science where similar basic principles are fundamental to models, but which might not be satisfied by the observations that are used for model evaluation.

Improving the Predictive Skill of a Distributed Hydrological Model by Calibration on Spatial Patterns With Multiple Satellite Data Sets

“…Additional key questions to address in this context include the model structural deficiencies (Gupta et al, 1998;Gupta et al, 2012) and the uncertainties of modeling data sets (i.e. input, calibration, and evaluation data), which can lead to erroneous model rejection (Beven, 2010(Beven, , 2018(Beven, , 2019b.…”

Developing observational methods to drive future hydrological science: Can we make a start as a community?

“…Hydrology is still, and for good reasons, an inexact science (for a recent discussion, see Beven, ), even if evolving hydrological understanding has provided a basis for improved water management for at least the last three millennia. The limitations of that understanding have, however, become much more apparent and important in the last century as the pressures of increasing populations, and the anthropogenic impacts on catchment forcing and responses, have intensified (see Abbott et al, ; Montanari et al, ; Sivapalan, Savenije, & Blöschl, ; Srinivasan et al, ; Wagener et al, ; Wilby, ).…”

“…This sophistication has created an illusion of real progress, but a case can be made that we are still rather muddling along, limited by the significant uncertainties in hydrological observations, knowledge of catchment characteristics, and related gaps in conceptual understanding, particularly of the subsurface. These knowledge gaps are illustrated by the fact that for many catchments, we cannot close the water balance without significant uncertainty (e.g., Beven, ; Schaller & Fan, ), uncertainty that is often neglected in evaluating models for practical applications. This lack of water balance closure can also result from a lack of information about the influence of water management on the water balance.…”

“…It might also allow much stronger testing of distributed model predictions in ways not previously possible, subject to having adequate observations of the input forcing. Despite the large and ongoing investment in rainfall radar methods and better rain gauges, a major limitation on model testing is knowing just what the inputs to a catchment area are within complex terrain (e.g., Beven, ; McMillan, Krueger, & Freer, ; Yatheendradas et al, ). If we did have better methods for estimating catchment inputs and discharges, we would be able to make much more rigorous hypothesis tests given information about storages and residence/transit times.…”

A process‐based approach to attribution of historical streamflow decline in a data‐scarce and human‐dominated watershed

“…Attribution in hydrology commonly follows a process of predictive inference, whereby models are used to infer causal relationships (Ferraro, Sanchirico, & Smith, ). In this approach, plausible mechanisms of change are incorporated into hydrological models (Beven, ), with the resulting models forming hypotheses about system behaviour (Beven, ), which can be rejected where they fail to capture relevant drivers of change (Beven & Lane, ). Predictive inference rests on the assumption that models which reasonably reproduce observed hydrological behaviour sufficiently represent the underlying hydrological processes (Ferraro et al, ).…”