A toolbox to evaluate data reliability for whole-ecosystem models: Application on the Bay of Biscay continental shelf food-web model

Abstract

Ecosystem models are always simplifications of reality and as such their application for ecosystem-based management requires standard validation. Here, the “DataReli” toolbox is proposed to evaluate the quality of the data used during the construction of ecosystem models, their coherence across trophic levels, and whether data limitations prevent the model long-term applications. This toolbox is the combination of three operational and complementary analyses: (i) the pedigree index to determine to what extent a model was calibrated on data of local origin; (ii) the graphical analysis known as PREBAL to assess whether a model respects some basic ecological and fisheries principles; and (iii) a sensitivity analysis to evaluate the robustness of model predictions to small variations in input data. The toolbox is delivered to potential users with main generic recommendations on how interpreting results conjointly and on which decisions to make about parameters’ revisions or model uses’ restrictions. (i) Corrections of parameters should be preferentially envisaged when modelling data-rich environments. (ii) For those models with an overall pedigree index above 0.4, a closer look at the pedigree routine, i.e. values by parameters and compartments, and the PREBAL analysis would help to prioritize parameters needing improvement. (ii)’ For Ecopath models of no overall acceptable quality (overall pedigree index <0.4), we recommend stopping the DataReli procedure at this point. (iii) In terms of sensitivity analysis, marked responses of model predictions to small variations in the input values must preferentially lead to restrictions in the model applications compared to corrections of parameter estimates. A concrete application of the “DataReli” toolbox to the pre-existing Ecopath model of the Bay of Biscay continental shelf food web is presented. For the present case study, the general level of input data reliability is considered as satisfying with regard to the model applications.