Dr Stephanie Brodie1, Dr Gemma Carroll1, Dr James Thorson3, Ellen Willis-Norton1, Dr Steven Bograd2, Dr Elliott Hazen2, Dr Kirstin Holsman3, Dr Jameal Samhouri4, Dr Melissa Haltuch4, Dr Stan Kotwicki3, Dr Rebecca Selden5
1University Of California Santa Cruz, Monterey, United States, 2NOAA Southwest Fisheries Science Centre, Monterey, United States, 3NOAA Alaska Fisheries Science Center, Seattle, United States, 4NOAA Northwest Fisheries Science Center, Seattle, United States, 5Rutgers University, New Brunswick, United States
In an era of extraordinary environmental variability and change, there is increasing need to accurately describe the patterns of species distribution and understand the processes shaping those patterns. Species distribution models (SDM) are now the most common approach to describing species distributions. Space and time covariates in SDMs can effectively describe species distribution patterns, but including environmental covariates can better address the processes involved. Not all SDM model types equally attribute environmental covariates and there is a need to explicitly explore if environmental covariates are realistically and accurately reflected in SDMs. To achieve this, we first used simulated data to build and compare three types of species distribution models of increasing complexity: a machine learning model, a semi-parametric model, and an autoregressive model. Secondly, we apply the same comparative model framework to a case study with three species (arrowtooth flounder, pacific cod, and walleye pollock) in the Eastern Bering Sea, USA. Spatiotemporal covariates effectively described species distributions, with the addition of environmental covariates generally improving model fit. In the models tested here, there was a trade-off between accurately estimating species abundance and accurately estimating the environmental mechanisms influencing species distributions. This trade-off between pattern and process highlights the importance of considering and identifying model purpose as a first step in the SDM process. Models used to transfer beyond the study domain (i.e. climate projection) need to focus on understanding processes, relative to models that are used to accurately describe the patterns in species abundance and distribution.
Steph is a post-doctoral researcher based in California. Her presentation focuses on considerations in species distribution modeling. Rest assured that all equations in the talk have been replaced with pretty pictures.