results of ais predictive modeling in a mn county...growing strong industries ~ developing new ideas...
TRANSCRIPT
Growing Strong Industries ~ Developing New Ideas ~ Nurturing Natural Resources
Josh DumkeKristi Nixon
https://data.nrri.umn.edu/ais/
Research guiding action: results of AIS predictive modeling in a MN County
St. Louis County Introduction Risk Assessment
Putting St. Louis County into Perspective
1. Connection to waters with most AIS in state: St. Louis River and Lake Superior
2. Lots of lakes; a rich network for secondary introductions because neighbors have lots of lakes too
3. Some well-known lakes, but many are ‘under the radar’ regarding AIS risk and prevention (132 lakes have boat access)
4. Is the recipient of the greatest AIS Prevention Aid due to high numbers of boat launches and parking spaces (>$710,000 for 2019)
St. Louis County Introduction Risk Assessment
Question: How to guide St. Louis County AIS Prevention Aid resources to have greatest impact preventing more invasions?
?MPR photo
https://whatcomboatinspections.com/
St. Louis County Introduction Risk Assessment
Introduction Risk Defined: the risk a lake has to receive AIS using 23 species from 2015 plan
We used models to identify lakes at greatest risk, but models need lots of data• Public AIS databases• Regional DNR offices• Spatial data downloads• Data from collaborators
Data – Boater SurveysBoaters surveys from MN DNR (2015-16) and 1854 Treaty Authority (2012-15) used to create a travel network and determine most frequent distance driven
Risk that AIS will remain alive as hitch-hikers diminishes as drying time of watercraft increases.
Data – Boater SurveysBoaters surveys from MN DNR (2015-16) and 1854 Treaty Authority (2012-15) used to create a travel network and determine most frequent distance driven
Risk that AIS will remain alive as hitch-hikers diminishes as drying time of watercraft increases.
~62% within 20 miles
AIS presence – AIS presence information for 23 species
Launches – Access type (trailer/carry-in), # ramps, docks, and parking spaces.
Recreation– Presence of popular game fish like Lake Trout, Walleye and Muskellunge, and average number of registered fishing tournaments.
Lake attributes - Max depth, surface area, shoreline distance, littoral area, and a calculated shoreline development index (SLD)
Proximity attributes – Land parcel type (residential, undeveloped, etc.), population density (ppl/acre), distance to towns, and road densities
Buffer analysis – Counts of individual AIS, launches, and lakes within 20 miles of every lake in the dataset (n=1,139; lakes >5 acres and assigned a Basin ID)
Data - Building the Dataset
73 boat launches w/in 20 mi of
East Vermilion
Data – AIS Summaries
Nearly 88% of lakes are AIS-free146 lakes have at least 1 of the 23 AIS considered
Most records are for invasive vegetation
Dissemination of Results via Web Tool
Main page featuresEssentially a data-dump of accesses, survey points, game fish, and individual AIS species
https://data.nrri.umn.edu/ais/
Dissemination of Results via Web Tool
Spatial Analysis page featuresDisplay of travel network from boater surveys, advanced filters, and access spatial queries
https://data.nrri.umn.edu/ais/
Dissemination of Results via Web Tool
Spatial Analysis page featuresAdvanced filter example• Conveys Vermilion gets lots of distant
visitors in the riskiest <24 hr period • Some prior lakes infested with AIS not
already in Vermilion
Lake Vermilion <24 hr travel network
Upper Red Lake: Starry Stonewort
Modeling - Methods
Data Screening: removed border waters, variables with missing data, “outlier” lakes, and highly correlated variables. Thinned to about 20 unique variables.
A Zero-inflated Poisson (ZIP) model was used to predict AIS Count.❖ Good at handling lots of zeros in a regression (88% of AIS count were 0)
A Random Forest model was used to identify which variables were most important in making correct AIS Presence/Absence predictions.
❖ Machine learning algorithm; randomly splits the dataset and randomly selects variables many times to ‘learn’ how to make correct classifications
Modeling - Results
ZIP model to predict AIS count (number of AIS expected)• More AIS predicted in 43 lakes than presently observed (R2 =59%) – These lakes would be good candidates for early detection surveys
Random Forest model to predict AIS presence (whether any AIS present)• 5.6% out-of-bag error
Modeling – Random Forest Results
Modeling – Random Forest Results
Assigning Relative Risk to St. Louis County Lakes
SumRank criteria for top 5 predictive variables1. Presence of a trailer boat launch2. 90th percentile for road distance w/in 30 m of shoreline3. 90th percentile for shoreline complexity index (SLD)4. Presence of either Walleye, Muskellunge, or Lake Trout5. 90th percentile for surrounding population density
❖ By ranking the entire dataset with these variables, we can assign AIS introduction risk categories to every lake in the County.
Dissemination of Results via Web Tool
Risk Assessment page featuresAIS introduction risk displayed for all lakes in the CountyCategories classified as Lowest Risk (0), Moderate Risk (1-3), and Highest Risk (4-5)
https://data.nrri.umn.edu/ais/
Dissemination of Results via Web Tool
Risk Assessment page featuresGuides for early detection activities – 43 lakes based on ZIP modelPlus a calcium query tool and flowlines
https://data.nrri.umn.edu/ais/
Summary
The Web Tool provides:
• …baseline risk of any AIS arriving, and places all lakes of St. Louis County on a relative scale to inform decisions
• …access to advanced filters to drill down into AIS, lake, and landscape data to help interpret AIS introduction risk
• Live version online now at https://data.nrri.umn.edu/ais/
Rusty Crayfish Presence
Additional Slides
Data - Building the Database
Data derived from (mostly) public sources:❖ Inclusion of collaborator survey data❖ Spatial data downloads (GIS layers and attributes,
water chemistry, land parcel designations)❖ MN DNR Lakefinder; infested waters lists❖ AIS presence websites (USGS NAS, MISIN,
GLANSIS, EDDMapS)❖ Inquires to MN DNR and SWCD regional offices
Resulting in:❖ 1,139 lakes >5 acres with a DOW❖ Comprehensive AIS presence information of
the 23 species listed in 2015 County Plan❖ 170 columns of information for each lake
(about 100 descriptive)
Rusty Crayfish Presence
Modeling – Random Forest Results
r = 0.65**
r = 0.40**
r = 0.37**r = 0.44**
r = 0.12**r = -0.19**
r = 0.12**
r = 0.23**
r = -0.23**
r = -0.06*
r = 0.08*
r = 0.14**r = 0.12**
Correlation of each variable to AIS Count (Pearson’s r)
Web Tool SummaryWhat the web tool doesn’t do Why?
It does not consider habitat suitability or probability of successful AIS establishment
Most lakes have no information on variables of ecological significance (e.g. [Ca], secchi, max depth, % littoral), and many AIS have unknown ecological thresholds
It does not make timeline predictions (e.g. 2 new AIS present by 2025)
Would require data on propagule loading, # of boats per season, and habitat suitability to make projections – beyond project scope
Not a predictor for individual AIS Web Tool assesses baseline risk of any AIS arriving, and places all lakes of St. Louis County on a relative scale.
Is not a new AIS mapping system Other databases are designed for conveying AIS sightings already (USGS NAS, MISIN, EddMapS, GLANSIS).
Live version online now at https://data.nrri.umn.edu/ais/
Collaborators – Boater Surveys
Watercraft Type:Majority of vessels recorded were classified as fishing boats ~17’ length regardless of use group
What this Web Tool Isn’t
• This does not consider habitat suitability or predict whether individual AIS will establish
• This web tool does not make timeline predictions (e.g. Zebra Mussel will be present by 2025)
• This is not a new AIS mapping system; there are already several databases that do that:
Dissemination of Results via Web Tool
Risk Assessment page featuresSLRE high-risk designation: meets all 5 risk-based criteria
Doesn’t the SLRE already have all the AIS it could get?? Well, no. It is still at high-risk of receiving new AIS (e.g. starry stonewort), due to all the risk factors (lots of people and lots of access options)
Pop-up for SLRE
Data – AIS Survey Data From CollaboratorsInvasive Plants-NRRI
Spiny Waterflea-USFS and 1854 TA
Rusty Crayfish-USFS and 1854 TA
Dissemination of Results via Web Tool
Risk Assessment page featuresFlowlines – consider lake position in a watershed; many are naturally connected
Dissemination of Results via Web Tool
Risk Assessment page featuresCalcium data – only available for some lakes, but could be useful in determining lake habitat suitability for zebra mussel. User-defined limits can be set with the query tool.
All Calcium (n = 75) Results for minimum limit of 20 mg/L (n = 13)
Data – Building Connectivity Buffers
There was no difference in distance between 24 hr and 1-4 day users traveling between lakes, so travel is pooled.
A 20 mile buffer captures over 60% of the connections between different lakes within 4 days.
Buffers used to collect information on lake surroundings