Bayesian Inference of Benthic Infauna Habitat Suitability along
Bayesian Inference of Benthic Infauna Habitat Suitability along the U. S. West Coast Andrea Havron M. S. Candidate, Marine Resource Management, Oregon State University Chris Goldfinger 1, Sarah Henkel 2, Bruce Marcot 3, Chris Romsos 1, Lisa Gilbane 4 1 Active Tectonic and Seafloor Mapping Lab, College of Earth, Ocean, and Atmospheric Sciences, Oregon State University 2 Benthic Ecology Lab, Hatfield Marine Science Center, Depart. of Zoology, Oregon State University 3 U. S. Forest Service, USDA 4 Bureau of Ocean Energy Management
Objectives • Development of offshore renewable energy • Limited knowledge of benthic infauna • Need to understand suitable habitat • Goals – Habitat Suitability Maps – Uncertainty Maps – Communicate use and limitation maps Oregon State University Conceptual Wave Park Wave Energy Anchor on Seafloor
Objectives • BOEM development of renewable energy • Limited knowledge of benthic infauna • Need to understand suitable habitat • Goals – Habitat Suitability Maps – Uncertainty Maps – Communicate use and limitation maps
Bayesian Networks • Benefits of Bayesian Approach – – Can handle missing data Remains robust with small datasets Easily incorporates multi-collinearity Easily tracks uncertainty P(B|A)P(A) P(A|B) = P(B) • Model Usability – Limitations of Benthic Infauna Models • Probability of suitable habitat • Static - based on data from summer 2010 -2012 • Focus on Physical Parameters – Reusability • Updateable with new information • Structure can be reapplied to new infauna species
Methods – The Data • Species Data – 218 benthic grab samples • Sternaspis fossor Axinopsida serricata Aystris gausapata • Habitat Data – – – – Percent Silt/Sand Total Nitrogen (TN) Total Organic Carbon (TOC) Mean Grain Size Latitude Depth Distance to Shore Local In Situ Data
Methods – The Data • Species Data – 218 benthic grab samples • Sternaspis fossor Axinopsida serricata Aystris gausapata • Habitat Data – – – – Percent Silt/Sand Total Nitrogen (TN) Total Organic Carbon (TOC) Mean Grain Size Latitude Depth Distance to Shore Local In Situ Data Regional Raster Data
Methods – Discretization density Absent Present Mean Grain Size Mean Absent Present • Discretize continuous variables • Breakpoints decided from field data
Methods – Model Structure • Basic Net – Arrows indicate correlation/causation – No multi-collinearity considered
Methods – Model Structure • Supervised Structure – Correlations from Field Data – Scientific Review
Methods – Model Structure • Intermediate nodes re-discretize regional raster variables to best predict local in situ variables
Methods – Model Structure
Methods – Training and Testing Models • Training – Inserted priors • Trained model with regional Grain Size database – Expectation Maximization Learning Algorithm • Calculates conditional probabilities • Can incorporate values with missing data • Testing – Performed 4 -fold cross validation – Evaluated performance metrics • Error Rates (0 -100%) • Spherical Payoff (0 -1) • True Skill Statistic (-1, 1)
Methods – Models to Maps • Model Selection • Prediction of habitat suitability based on four regional raster layers: – Depth – Mean Grain Size – Distance to Shore – Latitude • Uncertainty Maps – Measure of confidence in probabilities – Percent field data used to inform probabilities
Results – Sternaspis fossor
Results – Sternaspis fossor Train Error SP TSS Test Rate All Data 8 % 0. 94 0. 74 4 -fold cv 10 % 0. 91 0. 72 Probability of Habitat Suitability 0 - 0. 25 - 0. 49 - 0. 51 - 0. 75 - 1 Uncertainty Very Unlikely Somewhat Unlikely Completely Unknown Somewhat Likely Very Likely Experience Low: 0 High: 1 High: 0. 5 Low: 0
Results – Axinopsida serricata
Results – Axinopsida serricata Train Test All Data 4 -fold cv Error SP TSS Rate 11 % 0. 91 0. 58 15 % 0. 88 0. 55 Probability of Habitat Suitability 0 - 0. 25 - 0. 49 - 0. 51 - 0. 75 - 1 Uncertainty Very Unlikely Somewhat Unlikely Completely Unknown Somewhat Likely Very Likely Experience Low: 0 High: 1 High: 0. 5 Low: 0
Results – Aystris gausapata
Results – Aystris gausapata Train Test All Data 4 -fold cv Error SP TSS Rate 21 % 0. 85 0. 55 42 % 0. 68 0. 09 Probability of Habitat Suitability 0 - 0. 25 - 0. 49 - 0. 51 - 0. 75 - 1 Uncertainty Very Unlikely Somewhat Unlikely Completely Unknown Somewhat Likely Very Likely Experience Low: 0 High: 1 High: 0. 5 Low: 0
Translating Results to Managers • Communicating Uncertainty – Renewable energy development – Limited knowledge of benthic infauna – Extrapolating sample data to regional maps • Bayesian Network Models – Habitat Suitability – Uncertainty – Experience Ennucula tenuis Callinax pycna Magelona berkeleyi Aystris gausapata Onuphis iridescens Sternaspis fossor Axinopsida serricata
Thank you!
- Slides: 21