Effect of study area bathymetric heterogeneity on parameterization and performance of a depth-based geolocation model for demersal fishes

Abstract

State-space geolocation models can provide valuable information on the large-scale movements of many fish species. The sensitivity of such complex models to model assumptions and fixed parameters is rarely assessed quantitatively, yet is important for interpretation of results and adaptation for new species and different geographic regions. We hypothesized that parameterization and performance of a discrete Hidden Markov Model (HMM) with a Gaussian depth-based data likelihood for demersal fishes first implemented in the flat terrain of the North Sea would be affected by the more heterogeneous depths found in the North Pacific Ocean. We ran the HMM on depth data from simulated random walk movement trajectories in flat, sloping, and heterogeneous study areas in the North Pacific Ocean where known depth distributions in each model grid cell were provided by high-resolution (5 m) multibeam bathymetry data. Performance was compared among different data likelihood specifications and grid sizes in each area. We found that model performance decreased when grid cell depth distributions departed from normal distributions. Performance decreased with increasing grid size in the heterogeneous and sloping study areas but not the flat study area. A new method for specifying grid cell depth variance based on study area slope performed better than the standard method of obtaining variance from adjacent grid cell values for larger grid sizes in heterogeneous and sloping areas. Overall model performance was highest in the heterogeneous and sloping areas at small grid sizes and in the flat area at large grid sizes. The estimated value of diffusion was also sensitive to bathymetric heterogeneity and variance-specification method. These results suggest that the degree of study area heterogeneity should be considered when choosing fixed parameters such as likelihood and grid size, and when interpreting the model results. In addition, this approach demonstrates the need for sensitivity analyses when using the model on a new species and in a new study area.