12
YKL REA Northern Pike Model Photo: ADF&G

YKL REA Northern Pike Model Photo: ADF&G. Fish Distribution Models Photo: USFWS Evaluate model performance Classification tree and random forest models

Embed Size (px)

Citation preview

YKL REA Northern Pike Model

Photo: ADF&G

Fish Distribution Models

Photo: USFWS

Evaluate model performance

Classification tree and random forest

models

ADF&G AFFID species

occurrence data

GIS source data

Predict species habitat across REA

study area

Fish distributions

Create stream network and

landscape predictor variables in GIS

Process AFFID data for use in models

Stream Network Used TauDEM to process DEM1. Add in additional HUCs on boundary of study area that

flow into the study area2. Fill pits3. Calculate flow direction (D8 method)4. Calculate contributing area 5. Create stream network based on curvature method and

drop analysis

Predictor Variables

Photo: USFWS

Predictors of Fish HabitatElevationPermafrostGradientFloodplainSlope over area ratioStream orderWatershed areaAverage watershed annual precipitationAverage watershed annual temperatureAverage watershed elevationAverage watershed slope over area ratioAverage watershed slopePercent permafrost cover in watershedPercent lake cover in watershed

Process AFFID data- Presences from AFFID

and ADF&G/BLM telemetry project in Kuskokwim

- Absences from projects in AFFID that listed fish community sampling as an objective

- Resampled data in areas of high intensity (Pebble area and telemetry)

- Shifted points along flow direction grid until they reached the stream network

- Extracted all predictor variables to each data point

Classification Trees

Photo: USFWS

Classification Tree Analysis Steps:– Identify the groups– Choose the variables– Identify the split that

maximizes the homogeneity of the resulting groups

– Determine a stopping point for the tree

– Prune the tree using cross-validation

Absent0.97(263)

Asterospicularia laurae

Shelf: Inner, Mid Shelf: Outer

Absent0.78(64)

Location: Back, Flank Location: Front

Depth < 3m Depth ≥ 3m

(De'Ath and Fabricious 2000)

Absent0.56(9)

Present0.81(37)

Misclassification rates: Null = 15%, Model = 9%

Random Forests

Creates many classification trees and combines predictions from all of them:- Start with bootstrapped samples of data- Observations not included are called out-of-bag (OOB)- Fit a classification tree to each bootstrap sample, for each

node, use a subset of the predictor variables- Determine the predicted class for each observation based

on majority vote of OOB predictions- To determine variable importance, compare

misclassification rates for OOB observations using true and randomly permuted data for each predictor

Run models in Rct1<-mvpart(pres.f~.,data=fish.pred1[s1,],xv="1se")rf1<-randomForest(pres.f~.,data=fish.pred1[s1,],ntree=999)

Photo: USFWS

CT training CT validation RF training RF validation

1 0.096 0.161 0.108 0.113

2 0.108 0.194 0.092 0.161

3 0.12 0.161 0.096 0.097

4 0.12 0.145 0.116 0.129

5 0.108 0.194 0.108 0.145

6 0.072 0.097 0.112 0.048

7 0.124 0.177 0.108 0.097

8 0.112 0.097 0.104 0.081

9 0.137 0.081 0.124 0.065

10 0.12 0.145 0.141 0.097

summary 0.1117 0.1452 0.1109 0.1033

Model Performance

Photo: USFWS

Confusion Matrix0 1 Error

0 184 13 6.6%1 21 93 18.4%

Top five variables are watershed area, stream order, stream elevation, percent of watershed covered by lakes, and stream floodplain.

Northern Pike

Results:~ 10,900 km of predicted summer habitat (restricted to stream reaches > 1 km in length)

Predictor Presence AbsenceWatershed area 13,000 km2 60 km2

Stream elevation 60 m 200 mStream floodplain Yes NoWatershed lake cover 2.8% 2.1%

Stream order 4th 1st

Invasive MacrophytesClimate Change

Precipitation

Permafrost

FireHuman Uses

Mining

Infrastructure

Harvest

Contaminants

Temperature

Perm

afro

st

thaw

Reduction in age at maturity and shift in spawning season

Bio

accu

mul

atio

n of

m

ercu

ry in

adu

lts

Exp

ande

d ic

e-fr

ee s

easo

n

Tem

pora

ry in

crea

ses

in n

utri

ent i

nput

s

Elodea

ssp

coul

d re

duce

qua

lity

of s

paw

ning

ha

bita

t

In creased toxicity

Increased potential for establishment of invasive macrophytes and changing fire dynamics

Incr

ease

d co

ntam

inan

t so

urce

s

Cha

nge

in

depo

sitio

n ra

tes

Northern PikeEsox lucius

Habitat

Increase depth of active layer will increase lake drainage area

Subs

iste

nce

harv

est p

ress

ures

on

over

win

teri

ng p

opul

atio

ns

Dir

ect d

estr

uctio

n of

hab

itat,

hind

ranc

e of

mig

ratio

n ro

utes

, inc

reas

ed d

owns

trea

m tu

rbid

ity a

nd s

edim

enta

tion

Change AgentsDriversCEGeneral EffectCE-Specific Effect

Incr

ease

d w

inte

r

prec

ipita

tion

may

inc

reas

e ov

erw

inte

ring

hab

itat

Review

Please review and provide comments:- Distribution models for fish and habitats- Conceptual models and text descriptions for fish

Contact: Rebecca [email protected], 907-786-4965

Photo: USFWS