Extrapolation in species distribution modelling. Application to Southern Ocean marine species
In: Progress in Oceanography. Pergamon: Oxford,New York,. ISSN 0079-6611; e-ISSN 1873-4472, more |  |
| Keyword | | | Author keywords | Multivariate Environmental Similarity Surface (MESS) Marine species, Antarctic, Modelling relevance, Conservation issues |
| Abstract | Species distribution modelling (SDM) has been increasingly applied to Southern Ocean case studies over the past decades, to map the distribution of species and highlight environmental settings driving species distribution. Predictive models have been commonly used for conservation purposes and supporting the delineation of marine protected areas, but model predictions are rarely associated with extrapolation uncertainty maps.In this study, we used the Multivariate Environmental Similarity Surface (MESS) index to quantify model uncertainty associated to extrapolation. Considering the reference dataset of environmental conditions for which species presence-only records are modelled, extrapolation corresponds to the part of the projection area for which one environmental value at least falls outside of the reference dataset. Six abundant and common sea star species of marine benthic communities of the Southern Ocean were used as case studies. Results show that up to 78% of the projection area is extrapolation, i.e. beyond conditions used for model calibration. Restricting the projection space by the known species ecological requirements (e.g. maximal depth, upper temperature tolerance) and increasing the size of presence datasets were proved efficient to reduce the proportion of extrapolation areas. We estimate that multiplying sampling effort by 2 or 3-fold should help reduce the proportion of extrapolation areas down to 10% in the six studied species. Considering the unexpectedly high levels of extrapolation uncertainty measured in SDM predictions, we strongly recommend that studies report information related to the level of extrapolation. Waiting for improved datasets, adapting modelling methods and providing such uncertainy information in distribution modelling studies are a necessity to accurately interpret model outputs and their reliability. |
|
