Modelling approaches to distinguish whiting species in mixed-species commercial catches, and the impact on stock status metrics

Catch allocation models can split aggregated mixed-species catches into individual species for stock assessments and fisheries management. In this paper, we evaluate a suite of these models for splitting mixed ‘trawl whiting’ catches into eastern school whiting ( Sillago flindersi ) and stout whiting ( S. robusta ) allocations for a commercial ocean prawn trawl fishery in New South Wales (NSW), Australia. Accuracy of the models was evaluated against a scientific observer survey which accurately recorded species catches, and we compared the modelled allocations to an existing coarse ‘rule-based’ allocation. There was no single best structure for the allocation model, but the most successful models included depth as a covariate because this helped split species along habitat preferences. The model-based allocations reduced trip-level error by around 40 % compared to the existing rules, and removed an existing bias in total catch estimates. This led to altered time series of catches and catch-per-unit-effort, especially for northern zones. When data were analysed for the entire NSW region, the catch allocation process (existing or modelled) had little impact on resulting indices of relative abundance for each whiting species, even when using a spatio-temporal standardization model. This was likely due to changes affecting scale rather than trend and our indices being rescaled to better compare time periods. Therefore, past stock assessments relying on statewide indices derived from existing rule-based allocations are likely reliable. Nevertheless, the modelled allocations were more accurate at a local and zonal scale, which will enable analyses with a finer spatial resolution in future stock assessments. Additional observer surveys are an important tool for ongoing improvement and validation of our allocation models. • We evaluate a suite of catch allocation models for splitting mixed-species trawl catches and compare estimates to existing coarse allocations. • Models included binomial general additive models (GAMs), zero-one inflated beta regressions and a static habitat suitability ratio model. • The models were fit using a Bayesian framework to help propagate uncertainty from the allocation models to subsequent catch-rate standardisations. • Our species distribution modelling (SDM) ensemble approach with model-averaged outputs provides robust and defensible species-split estimates. • The modelled allocations were more accurate at a local and zonal scale, enabling analyses over finer spatial scales in future stock assessments.