Improving representation of zero flows in probabilistic hydrological modelling of ephemeral catchments

Ephemeral catchments, where streamflow is frequently zero or negligible, are common across the world yet difficult to model reliably. This paper evaluates probabilistic approaches for modelling streamflow in ephemeral catchments, with a focus on the description of predictive uncertainty using residual error models.

We compare an explicit treatment of zero flows using a censoring approach versus a simpler pragmatic approach where the lower streamflow bound of zero is applied in prediction only. Following a theoretical exposition, empirical comparisons are reported using a daily rainfall-runoff model (GR4J), four residual error schemes (based on log, log-sinh and Box-Cox (BC) transformations with power parameter L = 0.2 and 0.5), 74 Australian catchments with diverse hydroclimatology, and five performance metrics, including reliability, precision, bias and proportion of zero flow days.

The explicit approach is most beneficial in "mid-ephemeral" catchments (5-50% zero flows) where it offers substantial improvements over the pragmatic approach. The BC0.2 and BC0.5 transformations are Pareto optimal: BC0.2 achieves better characterisation of predictive uncertainty, whereas BC0.5 attains lower volumetric bias. In "low-ephemeral" catchments (<5% zero flows) the pragmatic approach is sufficient, whereas in "high-ephemeral" catchments (>50% zero flows) both approaches incur limitations and further method development is warranted. The findings provide guidance on improving probabilistic streamflow predictions in ephemeral catchments.