Skip to main content

    Using two international synoptic surveys to test the predictive performance of krill habitat models in the Scotia Sea

    Request Meeting Document
    Document Number:
    WG-EMM-2023/34
    Author(s):
    J. Freer, C. Liszka, S. Fielding, G. Tarling, S. Thorpe, S. Hill, B. Krafft and G. Macaulay
    Submitted By:
    Dr Simeon Hill (United Kingdom)
    Approved By:
    Dr Martin Collins (United Kingdom)
    Abstract

    Antarctic krill (Euphausia superba) is a key species within the Southern Ocean ecosystem and also the target of its largest commercial fishery. Statistical models that correlate krill biomass to environmental covariates can help understand the drivers of krill distribution, an important factor to consider in managing the fishery. The synoptic surveys of krill across the Scotia Sea in 2000 and 2019 are unique in their extensive spatial coverage of the Scotia Sea and intensive collection of acoustically derived krill density data within a narrow temporal window. These complementary surveys provide independent datasets with which to develop and evaluate habitat models.

    We assess how well sub-regional and regional scale krill distribution models fitted to year-2000 observations predict the krill densities observed during the 2019 survey. We also develop a novel krill distribution model fitted to year-2019 observations. R2and Root Mean Squared Error metrics indicate that, with few exceptions, models built using year-2000 data performed markedly worse in explaining the variance observed when applied to 2019 data. Both the environmental covariates (explanatory variables) and the functional forms of relationships between the explanatory variables and krill density differ markedly between surveys. In addition, models generally performed poorly at predicting high krill densities, even within the datasets used to derive them. Nonetheless, models of shelf regions maintained their predictive power and environmental relationships better than those of offshore regions. Based on these results we advocate the use of multi-year datasets to build distribution models, as these may help capture inter-annual variability necessary for robust predictions.