Linking Temporal and Spa1al Data Sets for Hierarchical Bayesian
Network Analysis and Predic1on of Delta Smelt
Popula1ons BJ Miller
Bob Oliver
The first of two presenta7ons
Our Purpose
• A different predic7ve model, Bayesian Network Analysis, for delta smelt (and similar problems)
• Recommenda7ons to improve sampling & rou7ne monitoring
• Preliminary results ranking factors important to larval-‐juvenile delta smelt
The Delta Smelt Problem
• Abundance declined by 2 orders of magnitude this century
• On St/Fed Endangered Species lists • Persistent record low levels • Many regression-‐based analyses • No predic7on models to inform management
The Fish & Wildlife Manager and the Bank President
Fish & Wildlife Manager
• Numbers of Delta Smelt
• Mul7ple factors
• Factors act hierarchically
• Iden7fy important factors
Bank President
• Probability of loan default
• Mul7ple factors
• Factors act hierarchically
• Iden7fy important factors
The Fish & Wildlife Manager and the Bank President
Fish & Wildlife Manager
• Numbers of Delta Smelt
• Mul7ple factors
• Factors act hierarchically
• Iden7fy important factors
• No solu1on—fix everything
that can be fixed
Bank President
• Probability of loan default
• Mul7ple factors
• Factors act hierarchically
• Iden7fy important factors
• Solved using Bayesian
methods—credit ra1ng
Method
• Ini7al focus on 20 mm survey (1995-‐2014) – Samples for important early life stages – Concurrent samples for zooplankton prey – Samples biweekly
• Iden7fied possibly important factors • Considered hierarchical influences • Divided habitat into zones • Allowed for 7me variance in rela7onships
How Does the Method Work?
• Experts collabora7vely draw the influence diagram: BUGSAT
• Organize data • Analyze influence diagram with data • Modify the influence diagram based on expert opinion or analysis results
• Repeat un7l sa7sfied
length of
spaw
ning
perio
d
starva7o
n
pred
a7on
air tem
perature
phyto-‐
plankton
N/P con
c.
Asian clam
turbidity
N & P input
Delta inflow
Simplified hierarchy delta smelt abundance
Delta
inflo
w
aqua7c vegeta7
on
dam con
struc7on
SWP-‐CV
P en
trainm
ent
turbidity
near
pumps (adu
lts)
Old-‐M
iddle
River fl
ow
X2 (juven
iles)
expo
rts
San
Joaquin
River fl
ow
expo
rts
Delta
inflo
w
lethal water
tempe
rature
air
tempe
rature
water te
mpe
rature
prey
density
turbidity
sedimen
t washo
ut
turbidity
pred
ator
abun
dance
contam
inant
effects
contam
inant
loading
Delta
inflo
w
power plant
entrainm
ent
diversion
% sm
elt n
ear p
lants
water te
mpe
rature
air tem
perature
Resid
ence 7me
Delta inflow
Conceptual Model Delta Smelt
Delta Smelt Resiliency Strategy
3 July 2016
Chart from “Delta Smelt Resiliency Strategy”
Influence Diagram
The Data Problem
Data Issues
• Missing data • Sample dates and loca7ons vary from survey to survey
• Sampling does not cover all important loca7ons
• Important factors not sampled well
>noma
Vacav)lli~ V«<I\ 1\IIVII
Salv•dor VIchy Spring• ) Union
.L @> @) 1 f
Napa \ I
Elmira l ~eeo Libeny Farms
(@)
~ ~aisie Maine Prairie
Mve Points
• 7\ ® • C3A • Jtand C3
""'r·· ~~ . ' I ~ '
;·
lmola
@) e 20 mm Survey Vineburg
ScheiiVIIIe ® @ 346
34S , pson
Green Volley RockVIlle ' Fairfield
@)
e Environmental Monltori"l ~"l e USGS Water Quality Monltorii1J
) /~ • {~~~~ I walnutGrt e I @ Ryde '\
Wa \er T \.on Lat.ng ®
I Epgle~ree !IIAA • @) t san Francisco Merazo Blll!llll"' Creston delia Bay N•tlllnal (@)
343. ""'; / ~ =N21132 \ ' 9-: @ @ 659. /-:. ) (. -_; Skaggslsland Amerlc•n r·'f \ 542.. • ..609 6Q / 711 'Isleton ?c MD6 ~A2• canyon .. . , 606Griizlylslan~ - ' ) •
..., ··~. l'£ildlrleAr•• •024 @ MD7i. - \ SanPablo .,
.,...., ® ''"""" ~ \ MD7&. '"' ® 341. " D7 "i~ @. ~. ® Refuge ®@ ( ' lulyBay • 610 6:54. . .919
• ..1.. 411 652 •l..' 7 7 •• ~ 34CJ'allejo Q,\1 " j • 6021501 D9 Montezuma ~51 • • 70 • \ l ,.r-'~
• 407 Ryer nd 519 • Collonov•lle D4 ':"D22o15 • I 812 026 e a15 ' ) MD 0 334 335 i 25....... • 02 5 • • _l06 I. • 10 v 418 6' Sui"' 4 2 513 70465Q;§ kj 906 ...
13 • • 336 " . • 411 si>QreArM0508
SZOA SSJ;,;Jf1Q3 D4A.&'
N%01:1 • &. N: Avon Pit1sbu• • 802 • ...._ \ •
@
woodo/1
S4nlla!IJo:lfa ~~ii 1>42 ""' ··- 7. . liii(..,DS • ••• • ,._11 ~& 0~·90;1 n~ ;D10~L ' •
12 320
m -, NZII04 ~..« .. - ,. ..-0 804 bu• l • -""'!"':' 902 · >.u' e' ~ .. ~\. l.~~.~~ 14. D41 He<cui<s~ (fJ V AntiockD Britn'ad Oakley L ~ • ~ 910 ......_, ~ t('
Pinole , Gl~ ' ra··.. 1 "rD •
leNea11 ijJ "'' ·~0 ~ Black ~ san<IH<tl -;:r•D~ • P8 . unvy,Ciub BeaclJ J' P
0
ncord Dl•mo•1d Knight&en .., J 914
4 • Sto< f!!' El Soboonte ' Mines , ..... ~ - - - ~ 'tt S~nPablo RfiO!lt$ Pleasa/t flrll cta)'lon Sth'i~nvlllr Slxt.J Orv·f~Sr . \iott Grills @ 1
rr.~PWOWo
S:::.n l=r:md.::~n
Wildcat Reg•qn-/l'jirk BrentWood , ' CM)'Im ..L ,.. I R<chmon\ Reqt0081P~ri • , , \.
V \th rr•to WalnutCteek CD D•~I)'!BoY ,I ~10 0 • ''
Marsh Creek 0 f pM..A. / I
-~ ~ I ~ t o.@ Mt Di~hlo ;. \ ~ ~ ~~m ,\ 0Jindn •oss•.-Berkeley I \
v AIO~ ~~ff~ '•, ~ ~. ~ I @) @ ~ Diablo Regional .J' , . . l J Moraga laHraml"'s · · Prewve, r--1'-4
~..1. I <' Elff•r,yvtll'ii C.nyonVal'·'"sta neg<Onal Danville B/ackhilwll ( p~P"1t:, J . ..._ ·~ l m, Pie~m<mt h ,. '' Wrldemes~ ' M,.... .,.., _
;;; v ~ ~ Cl" O•~•hd ,,_, ,...., -, ""-' J!,, ~ Q \. . " Houst •1,
Alamecta Sah Rlimort Is/arid
Conclusions
• Analyze with method that is appropriate for – Hierarchical influences – Varying rela7onships over years – Varying rela7onships by zones
• Lamina7on is necessary, but not ideal • Obvious requirements for rou7ne monitoring – Extend historical records – Sample where Delta Smelt are – Sample simultaneously for all poten7ally important factors
Conclusion
Bayesian Network Analysis – Is collabora7ve – Has been extensively used to solve important
problems – Requires sophis7cated, well-‐developed analy7cal
methods – Offers the possibility of conver7ng the hopelessly
complex problem to Delta Smelt into a simpler problem