Applications of Predictive Models for Invasive Plants Emma C. Underwood, Rob Klinger & James F. Quinn Environmental Science & Policy University of California - Davis
Why Use Predictive Models for Invasive Plants? Determine the likelihood of invasion over a broad scales Assist in setting priorities for control efforts Develop in conjunction with other mapping approaches
How Do They Work? Relate field observations to environmental predictor variables 1. Input data: presence, presence/absence, or abundance data 2. Environmental predictors: direct/indirect effects on species Limiting factors: e.g., temperature, precipitation, soil type Disturbances: e.g., natural, human related Resources: e.g., energy, nutrients, water
Process of Model Building 1. Conceptualization Natural history Model selection 2. Data preparation (data sources, scale) 3. Model fitting (correlation of variables) 4. Spatial predictions 5. Model evaluation 6. Assess model applicability (Adapted from Guisan & Zimmerman, 2000)
Model Selection Types of models Climatic envelope Classification & Regression Tree Genetic algorithms Generalized Additive Models General Linear Models Co-Kriging
Example 1. CART Iceplant: Vandenberg Air Force Base (Olmstead et al., 2004)
CART: Risk of Iceplant Invasion in Coastal Scrub Communities Coast Distance > 245m No Yes 76% 2 Soil Group Distance from Rivers < 406m 1 8% No Yes Elevation > 33m 14% No No Aspect < 109* No % Slope > 9.5 Yes Yes 0% Yes 29% 67% 37%
Iceplant mapped from imagery Iceplant risk of invasion (CART)
Model Evaluation Prediction errors Errors of omission Errors of commission Sources of errors Environmental Biological Algorithmic
(Underwood et al., 2004) Predicting Patterns of Non-Native Plant Invasions in Yosemite National Park, California
Yosemite National Park J. Randall. TNC Wildland Invasive Species
Objectives of Study Conduct community level analyses Develop landscape scale predictive model of invasive plants Suggest protocol for sampling of invasive plants in burned areas
Distribution of Invasives Determined by: I. Vulnerability of community to invasion II. Environmental niche of invasive species III. Areas of disturbance Natural (fire, flooding) Human related (hiking trails, campgrounds)
Fire Disturbance in Yosemite 1930-2000 Number of Fires 150 140 130 120 110 100 90 80 70 60 50 40 30 20 10 0 Fires permitted by NPS Burned areas are ideal environments for invasion 1920 1930 1940 1950 1960 Remove dominant species Increase bare ground, light, nutrients Decade 1970 1980 1990 2000 2010 Mean Burn Size (log ha) 4 3 2 1 0-1 -2 Yosemite 1930-2000 1920 1930 1940 1950 1960 Decade Y = 0.036(X) - 71.2 r = 0.66 P < 0.001 1970 1980 1990 2000 2010
I. Community Level Analyses Analysis of field data (N=236) Supplemented with GIS derived data Environmental variables: elevation, slope, aspect, % cover trees, shrubs, soil characteristics Disturbance variables: years since burn, size of burn, distance to road, trail, campground Regression analyses Ordination analyses Grouped co-occurring species
Identified key variables Elevation +1.0 Shrub Cover (%) -1.0 Slope +1.0 Tree Cover (%) -1.0
POAN AICA BRHO2 HYSCS2 PHPR3 TOAR BRRU2 PHAQ BRJA RUCR ERCI6 RUAC3 AGGI2 HYPE TRRE3 POCO DIAR LEVU TAOF MYDI POPR CIVU CILAL PONE LASE HOLA TRDU BRDI3 CECY2 GAPA5 VUMY CEGL2 BRTE BRAR3 AGST2 PLLA SIGA VIMA RUDI2 BRST2 AGCA5 5 4 3 2 1 0 Identified four species groups 1. Bromus tectorum, Vulpia myuros 2. Holcus lanatus, Lactuca serriola 3. Poa pratensis, Cirsium vulgare 4. Phyleum pretense, Hypericum scouleri 1 2 3 4 Ward's Distance Ward s Linkage
II. Predictive Modeling Goal To extrapolate from plot to landscape scale To compensate for limited field data Why GARP? Readily available No assumptions about underlying data Combination of approaches means greater predictive ability Novel application
Select plots with > invasive cover (80:20) Environmental Model Elevation Slope % Tree cover % Shrub cover Disturbance Model Distance from roads Distance from trails Distance from camps Predicted Distribution of Invasives
Model Evaluation Plots predicted as present Environmental Model = 76% Disturbance Model = 65% Environmental Model Verification 100% 1 75-100% 50-75% <50% Missed Disturbance Model Verification 100% 75-100% <50% Missed
Significance of Study Predictive model developed with an ecological basis Includes both environmental and disturbance variables Results provide foundation for NPS sampling and monitoring activities
Limitations of Predictive Models Risk of over- and under- prediction Static & fail to reflect environmental variability Models based on limited input data, extrapolation must be done carefully Temporal scale
Conclusions Models offer valuable tools for extrapolating to broader scales Multiple models available, allows flexibility Predictive modeling field is maturing, but requires shared experiences
Acknowledgements Peggy Moore; Western Ecological Research Center, USGS Linda Mutch; NPS Inventory & Monitoring Coordinator Marcel Rejmanek; University of California, Davis John Randall; Invasive Species Initiative, The Nature Conservancy National Park Service & Yosemite National Park David Stockwell; University of California, San Diego Karen Olmstead; CSTARS, University of California, Davis