2012 Pre-Season Forecasts for the Stillaguamish River Chinook EMPAR (Environmental Model Predicting Adult Returns) January 14, 2012 Developed By: Jason Hall 1 and Dr. Correigh Greene 2 1 HALL AND ASSOCIATES CONSULTING, INC. [email protected]2 NOAA NORTHWEST FISHERIES SCIENCE CENTER [email protected]
31
Embed
2012 Pre-Season Forecasts for the Stillaguamish River Chinook
2012 Pre-Season Forecasts for the Stillaguamish River Chinook. EMPAR ( E nvironmental M odel P redicting A dult R eturns) . January 14, 2012 Developed By: Jason Hall 1 and Dr . Correigh Greene 2 1 Hall and Associates Consulting, Inc . [email protected] - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
2012 Pre-Season Forecasts for the Stillaguamish River Chinook
EMPAR(Environmental Model Predicting Adult Returns)
January 14, 2012
Developed By: Jason Hall1 and Dr. Correigh Greene2
• Return rates are driven by survival across multiple life stages– Unique environmental conditions are experienced during each life stage– Life stage specific environmental conditions can influence survival
• Forecasts that consider life-stage specific environmental conditions may provide better forecasts– EMPAR developed to provide an accurate and robust forecast model
that incorporates life-stage specific environmental conditions– Approach adapted from Greene et al. (2005)*– EMPAR development started with 2009 return year forecast
*Greene, C.M., D.W. Jensen, G.R. Pess, and E.A. Steel. 2005. Effects of environmental conditions during stream, estuary, and ocean residency on Chinook salmon return rates in the Skagit River, Washington. Transactions of the American Fisheries Society 134:1562-1581.
3Jun-
00
Jan-
01
Jul-0
1
Feb-
02
Sep-
02
Mar
-03
Oct
-03
Apr
-04
Nov
-04
May
-05
Dec
-05
EGGHATCH
PINK/CHUMQMAX
Delta
Near
Ocean1
Ocean2
Ocean3
Ocean4
FW
Age 3 Spawners
Age 4 Spawners
Age 5 Spawners
DOTEMPSAL
SSTPDOSOIUWISL
Age 2 Spawners
Background: Life stage concept
4
Background: Broodyear Model Concept
RY 2000Spawners (S)
2002 2003 2004 2005 2006 2007 2008
RY 2001Spawners (S)
RY 2002Spawners (S)
RY 2003Spawners (S)
Age 2Age 3Age 4Age 5
SPSt = spawners per spawner in year tNt = adult escapement in year tPx,t = proportion of age x in return year t
5
RY 2000Spawners (S)
2002 2003 2004 2005 2006 2007 2008
RY 2001Spawners (S)
RY 2002Spawners (S)
RY 2003Spawners (S)
Age 2Age 3Age 4Age 5
*Return rate calculated for each age class – Age 3 example shown here
Background: Age-Specific Model Concept
6
Background: EMPAR Updates
• Removed some environmental factors from consideration:– Infrequent data update schedule– Forecast years rely on estimated data
• Added 2009 and 2010 return years to model training set: – Increases sample size by almost 10%– Return year 2011 was used as sole test set
• Working with age-specific models only: – Removes errors associated with applying average age structure– Makes more sense from a biological standpoint
7
• Incorporated Principle Components Analysis (PCA):– Common factor analysis technique– Synthesize multiple factors within a life stage into primary components – Longer temporal patterns can be considered – More arbitrary than using actual factors, but is more robust– Allows trends in many factors within a life stage to be considered
Background: EMPAR Updates
8
PCA Approach:• PCA for Freshwater Life Stage (1989-2010)
– EGG, PKCM, HATCH, and QMAX– 62% of variance explained with first two components
• PCA for Delta/Nearshore Life Stage (1989-2010)– DO, TEMP, and SAL– 50% of variance explained with first two components
• PCA for Ocean Life Stage (1949-2010)– SST, UWI, PDO, SOI, and SL – 73% of variance explained with first two components
9
PCA Approach:• Linear regression models:
– Combination of PCA components and selected raw factors– 2 freshwater, 1 delta/near, and 2 ocean life stage factors– PCA components (representing multiple factors) count as 1 factor
• Over-parameterized model?– Significant increase in predictive power for key age groups– Describes complicated life cycle well– Several evaluation techniques indicate that these models are not over-
parameterized
10
Test Set Validation: SNOR Models
1994 1996 1998 2000 2002 2004 2006 2008 20100E+00
1E+05
2E+05
3E+05
4E+05
5E+05
6E+05
Return Year of First True Forecast
Mea
n Sq
uare
Err
or
*MSE decays as test set increases
*Similar patterns observed for factor coefficients
11
Results: PCA EMPAR Model SummariesPopulation Age R2 P-Value F-Statistic DF Factor1* Factor2* Factor3* Factor4* Factor5*
* Factors in parentheses have negative coefficients** Pearson’s correlation calculated based on sum of predicted returns of each age class by return year and observed escapement
• With return years 2009 – 2011 as test sets:– Derived forecast from all three selected EMPAR models– PCA model shows best track record when compared on
equal terms
16
• Use EMPAR models that incorporate Principle Components Analysis (PCA):– Forecast trends track well with observed trends– Factor sensitivity does not appear to be an issue as compared to
the full permutation AIC based approaches– Forecast performance comparisons indicate that the PCA model has
better predictive accuracy– PCA model does not appear to be over-parameterized and training set
appears valid– More arbitrary than using actual factors, but is a more statistically
robust procedure– Allows consideration of trends in multiple factors within a life stage
Recommendations:
17
Results: 2012 Forecast
Age 2 Age 3 Age 4 Age 5 Total
SNOR 28 126 180 5 338
FNOR 1 6 77 1 86
SHOR 79 136 325 41 580
Age 2 Age 3 Age 4 Age 5 Total
SNOR 40 179 256 7 481
FNOR 2 9 110 2 122
SHOR 112 193 463 58 827
Escapement with Fishing
Escapement without Fishing (assumes average exploitation rates)
18
Results: 2012 Forecast FRAM Conversion
Age 2 Age 3 Age 4 Age 5 Total
SNOR 738 592 198 5 1534
FNOR 26 28 85 1 140
SHOR 2083 639 357 41 3121
Age 2 Age 3 Age 4 Age 5 Total
SNOR 1230 846 247 6 2330
FNOR 44 40 106 1 191
SHOR 3472 913 447 46 4878
FRAM Input MM Run
FRAM Recruits
19
EMPAR Supporting Information:
The following slides are supplemental information to support the presentation and detailed questions…
20
PCA Example: Delta/Nearshore Life Stage PCA
Component Variance Explained
Cumulative Variance
Delta DO
Delta SAL
Delta TEMP
Near DO
Near SAL
Near TEMP
1 0.27 0.27 - + - - +
2 0.23 0.50 - - -
3 0.20 0.70 + - - + - -
4 0.14 0.84 + - - + -
5 0.08 0.92 - - + +
6 0.08 1.00 - - + + -
21
PCA Example: Ocean Life Stage PCA
Component Variance Explained
Cumulative Variance SOI SL SST UWI PDO
1 0.49 0.49 + - - + -
2 0.24 0.73 - + - - -
3 0.15 0.88 - + +
4 0.07 0.96 - - - - +
5 0.04 1.00 + + - + +
0 100 200 300 400 500 6000
200
400
600
800
1000
1200
f(x) = 1.08714300474013 x + 65.2866563127376R² = 0.370966047733079
f(x) = 1.3834344925073 x + 92.0374287903067R² = 0.599812688008294
f(x) = 1.28117172740322 x + 91.7906305117302R² = 0.608144549939346
• Broodyear Model:– Simplest model structure – Calculate return rates for each broodyear– One model for all spawners produced from each broodyear– Separate model for SNOR, SHOR, and FNOR– Allocate predicted returns by average age structure
• Age-Specific Model:– More complicated model structure– Calculate return rates for each age class by broodyear– Separate model for each age class– Separate model for SNOR, SHOR, and FNOR
25
Background: Life stage factors • Freshwater (Aug – Feb)
– Egg deposition– Pink and Chum escapement – Hatchery Releases– Max incubation flow– Min spawning flow
• Ocean Year 1 – 4 (Oct – Sep)– Sea Surface Temperature– Upwelling Index – Pacific Decadal Oscillation– Southern Oscillation Index– Sea Level– Aleutian Low Pressure Index*– SVI boreal copepod*– SVI southern copepod*
*Removed from candidate list
26
Background: Life stage factors
27
Background: EMPAR Approaches
• Several model selection and model development approaches have been considered during the development of EMPAR: – Full permutation models with Akaike's Information
Criterion score (AICc) model selection– Stepwise regression techniques– Principle Components Analysis (PCA) based approach
28
EMPAR Approaches: AIC Models
• Full permutation models with Akaike's Information Criterion score (AICc) model selection – Multiple models provide information about dependent variables– The best models are those that have strong predictive power
but use fewer independent variables– AIC scores models based on their ability to reduce uncertainty
but penalizes by the number of variables in the model– Not sensitive to the order variables enter as in stepwise regressions
• Model structure caveats – Large test model sets increases risk of selecting randomly
correlated models– Sensitivity to collinearities were initially a problem, but were
subsequently resolved in later models– Forecast outputs show sensitivity to variations in strong factors,
but were more accurate than stepwise regression models
29
EMPAR Approaches: Stepwise Regression
• Stepwise regression model selection– Common and well established approach– An aggressive fitting technique that can be overly greedy
• Model structure caveats – Sensitive to factor order– Favors models with fewer factors, and therefore does not
consider all life stages – Stepwise regression approaches appear to produce less accurate
forecasts despite the caveats associated with the full permutation AIC approach
30
EMPAR Approaches: PCA
• Principle Components Analysis– Common factor analysis technique– Reduces the number of variables and detects structure within
a set of factors– Can be used to synthesize multiple factors within a life stage
into primary components – Longer temporal patterns can be considered since components
can be derived independently• Model Structure Caveats
– Models using PCA components can be more conservative– Interpretation of the influence of factors within components
is not as direct as in AIC or stepwise regression techniques