Title: Fish OE Modeling
1Fish O/E Modeling
- Aquatic Life/Nutrient Workgroup
- August 11, 2008
2Discussion Topics
- Reference site data
- Evaluation of fish O/E indices for speciose
streams - Initial site classification and predictive
modeling - Individual species models as an alternative
management tool for species of interest/concern - Continuing efforts
3Reference Site Data
- Data from 182 reference sites
- 151 sites from CO Division of Wildlife
- Sites from EMAP-West
- 4 samples contained 0 fish
- 36 native species used
- All trout considered native or desirable
- All cutthroats lumped in cutthroat group
4Reference Site Map
5Evaluation of O/E Indices
- Classify streams based on taxa composition
- What streams are similar biologically?
- Model biotic-environment relationships
- Usage of predictor variables
- Use model to estimate site-specific, individual
species probabilities of capture (pc) - E (expected), the number of species predicted at
a site Spc - Compare O (observed) to E to determine impairment
6Initial Classification of Reference Sites
- Composition of native or desirable fish species
at reference sites only - Biologically similar sites being grouped together
- Cluster analysis/ordination revealed several
relatively distinct groupings of sites based on
species composition - 10 classes selected
7Cluster Analysis Dendrogram
WHS, CRC, CSH, JOD, ORD, LGS, IOD, PTM, BMS
FHC, BBH, RDS, LND, SMM, CCF, SNF, BBF
PKF, FMW, STR, SAH, BMW, BST, ARD
- 9 classes (or species groups) based on species
composition - Indicator spp BHS, SPD, TRT, WHS, FHC, PKF (no
CPM)
8- Classes mapped by indicator spp
9Modeling Biotic-Environmental Relationships
Variables extracted from 403 samples
Product from Classifications
Cont.
10Model Prediction Errors w/ Trout
- No model is completely precise nor accurate
errors must be quantified - Trout (TRT) predicted correctly 93 of the time
- Bluehead sucker (BHS) wants to predict as TRT
or SPD ? 100 error
11Affects From Introduced Trout
Trout Thermal Limits (17.5 o C)
Source Utah State Univ.
- SPD and BHS groups are vulnerable to introduced
trout WHS slightly less vulnerable - Trout presence has muddled predictions in the West
12Model Prediction Errors w/o Trout
- Overall, predictions improve w/o trout
- BHS error drops to 31
13Estimating Probability of Capture
- Discriminant model output use to estimate E
- Sum PC (probability of capture)
- Probability of capture still a quantitative way
of predicting spp in individual spp modeling
14Initial Modeling Results
- A single, statewide model attempted
- Most speciose group has about 6 taxa per sample
on average, too few for precise O/E indices - Results indicate that model too course
Max 13
15Initial Modeling Outcome
- Failure to detect 1 spp could result in extensive
deviation in O E assemblages, which results in
low sensitivity - Not useful in a regulatory-sense
- WQCD took a shot at developing a practical
bioassessment tool for fish to complement
macroinvertebrate tools - Next step decompose model into individual taxa
models (species modeling)
16Benefits of Individual Species Modeling
- Predicted list of fish species
- Best estimate of historical distribution
- Antidegradation for high quality sites
- Visual tool (when predictions wired into stream
layer) - Statewide application
- Alleviates mountains issue
17Individual Species Modeling
18Model Types Used
- MaxEnt (Maximum Entropy) only uses presence
data - RF (Random Forest) uses observations from
both presence and absence data - Approach not based on normal classification and
regression tree (CART) work more like
bootstrapping
19Model Results
- Values range from 0 to 1
- 1 perfect model
- Many models above 0.8 ? should see good
predictions
AUC Area Under Operator Receiver Curve
20Model Results
- Those potentially affected by trout
introductions BHS, SPD WHS (indicator spp)
MTS (which groups w/ BHS)
AUC Area Under Operator Receiver Curve
21Applicability
- Can use this type of mapping for all 18 spp
- Probability (of capture) of finding that spp
wired into each pixel
22Ongoing Work
- 13 additional reference sites added to modeling
in July 08 (emphasis on plains and San Luis V.) - Will attempt using Similarity Coefficients
- 2 samples are x similar to ea. other
- Will attempt a John Van Sickle (EPA) Similarity
Index approach - How similar is O to E?
- Niche modeling i.e. where spp should be
23Summary
- Traditional RIVPACS modeling approach did NOT
work model not bad, just too course - Alternative approaches explored
- Individual spp modeling best performing approach
- Demonstrates strong utility in regulatory
framework - Modeling moving forward towards completion
24Questions?
Oncorhynchus clarki stomias
Catostomus discobolus
Cottus bairdii