Friday, April 15
10:15 a.m-11:45 a.m.
Session 10:
Predictive Modeling and Forecasting
182
-------
f*"!
Day Three: Session 10
J
iL2a
Implementing Predictive Models:
Practical Advice and New Tools
Adam Mednick, PhD
University of Wisconsin Sea Grant Institute
Abstract
Over the past 5 years, the practice of devel-
oping and implementing predictive models
at coastal beaches has increased several-fold,
particularly in the Great Lakes. During the first
3 years of the Great Lakes Restoration Initiative
(GLRI), the adoption of the U.S. Environmental
Protection Agency's (EPA's) Virtual Beach
decision-support software, among other
tools for implementing operational nowcasts,
expanded from a handful of sites to over 50
beaches. Whether this expansion will continue
in the absence of centralized model-building
services previously supported by GLRI remains
to be seen. At issue is whether a typical local
government (e.g., public health or parks depart-
ment) can develop, operate, and/or maintain
nowcast models without additional funding
or specialized staff. Based on past experience
and research, the presenter will argue that the
answer is a conditional "yes" and will provide
practical suggestions on how EPA and its state,
local, and academic partners can overcome both
real and perceived barriers, such as the lack of
adequate data, technical know-how, clear deci-
sion criteria, managerial confidence, and time.
The presentation will highlight issues relevant
to marine beaches, where adoption to date
has been minimal, and will conclude with an
updated look at the suite of resources and tools
being developed to make the process easier and
more sustainable over time.
Biosketch
Dr. Adam Mednick is a postdoctoral fellow
at the University of Wisconsin (UW) Sea Grant
Institute. He received his bachelor of science
degree in natural resources from the University
of Minnesota, his master of forest science degree
from Yale University, and his doctorate in urban
and regional planning from UW-Madison.
Dr. Mednick has worked in conservation policy
and planning, spatial analysis, research, out-
reach, and education on a range of issues at the
state, local, and national levels. Prior to join-
ing UW Sea Grant in 2014, he worked for the
National Parks and Conservation Association in
Washington, DC; the New Jersey Conservation
Foundation in Far Hills, New Jersey; and the
Wisconsin Department of Natural Resources
in Madison. Dr. Mednick is an elected member
the Great Lakes Beach Association board of
directors, a founding cochair of the Wisconsin
Coastal Beaches Workgroup, and the manager
of the Virtual Beach Users' Group. His cur-
rent professional interests include how best to
develop and deploy environmental data and
modeling systems to the benefit of real-world
decision making; and, more generally, how to
make academic and government research more
useful through collaboration and cooperative
extension.
183
-------
jl3
U.S. EPA's 2016 Recreational Waters Conference
uNMRsny of Wisconsin sea gmkt institiitt
Why Predictive Models (Nowcasts)?
2. To Increase the Frequency of Monitoring
Why Predictive Models (Nowcasts)?
1. To Reduce Unnecessary & Missed Advisories
Sampled Advisories:
Sampled Open;
Sampled All:
All
In-Error
All
In-Error
All
Total Errors
Illinois
1,107
709 (64%)
9,127
1,142(13%)
10,234
1,851 (18%)
Indiana
716
452 (63%)
3,881
555(14%)
4,597
1,007(22%)
Michigan
201
137(68%)
7,020
528 (8%)
7,221
665 (9%)
Minnesota
74
58 (78%)
1,853
90 (5%)
1,927
148 (8%)
New York
467
260 (56%)
2,322
440(19%)
2,789
700 (25%)
Ohio
636
434 (68%)
4,842
770(16%)
5,478
1,204 (22%)
Pennsylvania
72
59 (82%)
937
88 (9%)
1,009
147(15%)
Wisconsin
1,775
849 (69%)
9,868
951 (10%)
11,093
1.800(16%)
Total
4,498
2,958 (66%)
39,850
4.564(11%)
44,348
7,522(17%)
Data from EPA BEACON (2008-'12)
Communities using Nowcasts
25 Communities using 'Virtual Beach' (VB)
[a] Case Study Communities
1
¦ Communities using Custom Modeling Systems
¦ Communities using VB + Custom Systems
"f
.l.lflli
a
2003 2006 2007 20OS 2O09 2010 2011 2012 2013
Case Study Communities
Community A (The "Innovator" - adopted 2009)
"The intent [is to] rely on Virtual Beach more fully, so we
don't spend as much time and money on testing."
Community B (The "Early Adopter" - adopted 2011)
"We [experimented with] it year-to-year [to see] how well it
performed... how many Limes it [was] right or wrong..."
Community C (The "Early Majority" - adopted 2013)
"I went to my Administration and said... 'There's some guy
from Madison who's really pushing iL and I've heard
Ianother community] is using it'..."
From Mednick (201<1)
Theoretical "Diffusion" of
Nowcast Models
Based on Rogers (1962)
//
"Take-off"
/
/ /
\
(3.) "Early Adopters"
(1.) Innovator** jL
y M
(a) \
"fcarly
Late (5.) N.
, Majority"
t Majority* v
Implementing Predictive Models
Practical Advice and New Tools
Adam MeJiiitk — Uniwtsily uT Wistuiisiii Sed Gidiil lnslilule
EPA Rcorcational Waters Conference (New Orleans, April 15, 2016)
EE! WIVONSIN CfWTAI
11—3 Management Program ,
WisconsinIii.
184
-------
Day Three: Session 10
Critical Question:
Can a typical local health
department develop, operate,
and maintain its own
predictive model...
HI Without special funding?
13 Without spec.iHli7ed staff?
- Mednick & Watermolen (2014)
Expanded Beach "Nowcaet"
Modeling across Wisconsin
'Typical' Health Departments
From Rockwell et al. (2014):
• In nearly 40% of reporting depts,
interns responsible for over half
the beach-related work.
• Among depts responsible for > 5
beaches, over 75% devole < 10% of
overall time to beach-related work.
• Among primary staff responsible for
beaches, over 85% spend < 1/4 of
their time on beach-related work.
Other Perceived Barriers:
Lack of Data
• 86% of beach managers said
location-specific, Web-accessible
data would be 'very useful' or
'extremely useful' (tfl out of 5)
Lack of Tools
• 86%... said improved predictive
modeling tools would be 'helpful'
or 'very helpful' (#1 out of 9)
WfBr
i «-
h. x &
"Virtual Beach"
•j&WSe&ieL
1 GJobjl Dateihcct IM.J-, MIR
www.seagrant.wisc.edu/virtualbeach
Virtual Beach
'Virtual Beach' is fr« dcctaon ronton software for wisdom thai enables coastal
hearh managers, pnhlie health professionals, ami applied researchers to efficiently
- Wbethor to issue (or fitt) swim adwories; closures on a ffivnn dav
About Virtual Beach
Unlortunately. the types ol
nrionriye remediation efforts an* Often imrinvly;
Moii a II .ii i Hilton
fr
www.seagrant.wisc.edu/virtualbeach
< Back ta Virtqal Bw-h Imam Meel
Online Dntn
potentially re
such as stream discharge, antecedent rainfall,
wave height waterfowl etc... Depending on
the beach, mam- if not all of these data will be
available online, tor free
Water Onalltv ft Reach Condition*
Historical water quality can be downloaded
from EP.Vs BEACON website - or from the
; lisiAd hriow. Dejsendin j
adnliar) uuve) ilaM tor jour beadi. It b to
yost advantage to upload them to jour stale's
beach website. This will make building and
1.) Lab Data + Sanitary Conditions
2.) Hydro-Meteorological Data
i -
ievraUe tables of L. coH data. Single-day sumi
ii VlenaUte laWes tAt-mO data. Single-ilaj sum
ta single-day su
185
-------
U.S. EPA's 2016 Recreational Waters Conference
UNMRSfTY OF WISCONSIN SEA MAKTHCTTUTf
msGS
EffTTHT
'EnDDaT"
Actc^-j iind Inlctiralc Enviiumnciilul Ob&civuUui» for Cao-jlol Dcdwuu Supp
¦ I — I
585-713 43 890
-87.890
2002-01-01 201
iSMU
07.006
;
3M5M0'.»'
584:714 43.841
=>»?43.S79
-87.654
2002-01-01 201
NOAA Operational Forecast Systems
Phased Conversion to rinite Volume Coastal Ocean Mode! (fVCOM)
Fiuiri ZliddK (2013)
Northern Gulf of Mexico OFS Currents Nowcast
VnClrf <1* O'AO (CM) (U/lM/lifi
TimefDaw. ,j 1 j i y ^ ;0 v prey Slop Animauon | Nexi |
"EnDDaT" is National
Access and Integrate Environmental Observations few Coastal Decision Support
:h cross Dili I Create Ptojetl Lo
I EnDDaT Informiuon I
• weat Lakes coastal t-orecastina system biii-bi ~ siOKti Qt: water uuaiitv1
• USGS QW (Water Quality)
^arfh within a mil* hnunHing ho* frnm project (dick marker tft Identify)
www.seagrant.wisc.edu/virtualbeach
Virtual Beach
'Virtual Beach ilSrw decuioc fu
M m3-ac^lSro?T5ITP?S
About Virtual Beach
U^DloaO «r
hpwflts. -ilnng me i.v uresi i altes nun. an esrtmarwi n million
people visit beaches each rear. Unfortunately, the types ot
mmmimifare heatlfi mis and prionma remediation efforts are often imrimtly.
inacmmir and/or factimnfrfe Virtual Reach ATO addresses ihw
-------
Day Three: Session 10
Virtual Beach 3.0.6
1. 'Gradient Boosting Machine' (GBM) Method
- More efficient model-building (regression trees)
2. Direct connection
to 'EnDDaT'
BIG. Easy data:
River Discharge,
Waves, Currents, etc.
(Spatiotemporally-
matched/processed)
15. * <9
Traditional (MLR) Model-building
Vra\Wortih<«> Prowl.G6M.vUp
Other Perceived Barriers:
Limited Technical Know-How
• Over 60% said training on predictive
models would be 'helpful' or 'very
helpful' (Rockwell et al. 2014)
Lack of Comprehensive Guidelines/
"Best Practices"
• Under Development (UW Sea Grant)
Lack of Confidence on the part of
Administrators and Decision-Makers
*
www.seagrant.wisc.edu/virtualbeach
Virtual Beach
About Virtual Beach
¦ iflj U fc X ,•» 0
« * F I? £ C » - :
is Learning Modula II: Sactionc A - E
fa
KJOQiA fr
187
-------
Virtual Beach Users' Group
. M
Virtual Beach Usors
Best Practices (DRAFT)
Model Operation
SDaily, preferably between 8:30-10:30 am kDT
SOperate in conjunction with regular data reporting
¦S Report 'Model' as the reason for beach actions
Minimum Field Data (required in Wisconsin)
^'Clarity' (categories) ^Cloud Cover (categories)
t Turbidity (NTUs/Secchi cm) / Algae in water (categories)
•f Water Temperature
-------
f*"!
Day Three: Session 10
J
iL2a
California Beach Water Quality
Nowcasting
Leslie Griffin
Heal the Bay
Abstract
Traditional beach management that uses
concentrations of cultivatable fecal indicator
bacteria (FIB) may lead to delayed notification
of unsafe swimming conditions. Predictive,
"Nowcast" models of beach water quality may
help reduce beach management errors and
enhance protection of public health. This study
compared the performances of five different
types of statistical, data-driven predictive mod-
els—multiple linear regression model, binary
logistic regression model, partial least-squares
regression model, artificial neural network, and
classification tree—in predicting health advi-
sories due to FIB contamination at 25 beaches
along the California coastline. In total, over
700 models were developed and evaluated.
Multiple linear regression with threshold tun-
ing performed well, along with binary logistic
regression with threshold tuning and classifica-
tion trees. On average, models outperformed the
current method based on day-old FIB concentra-
tions by capturing 25% more poor water quality
days while maintaining equivalent false nega-
tive results. Beaches with well-performing mod-
els usually have a rainfall/flow-related dominat-
ing factor affecting beach water quality, while
beaches having a deteriorating water quality
trend or low FIB exceedance rates are less likely
to have a well-performing model. Based on the
results of this study, we carried out a pilot study
at three Californian beaches with beach man-
agers in the summer of 2015 to use daily now-
casting for public notification of beach water
quality. Due to the success of the pilot program,
the State of California has funded the develop-
ment of a Nowcasting system to provide daily
information to local beach managers in an effort
to help inform public notification decisions for
up to 25 separate beach locations over the next
3 years.
Biosketch
Ms. Leslie Griffin is the beach water
quality scientist at the Los Angeles-based
environmental organization, Heal the Bay.
Native to the East Coast, she relocated across
country to receive her bachelor and master
of science degrees in environmental science
with an emphasis in water quality from Loyola
Marymount University. She worked on passive
sampling of PAHs for 2 years while obtaining
her master's degree. While pursuing her educa-
tion, Ms. Griffin interned at Heal the Bay as an
aquarist and a watershed educator. In 2015, she
began working full time with the organization
as the data analyst for the Beach Report Card
program. Currently, Ms. Griffin manages the
Beach Report Card program—working to ensure
accurate and timely dissemination of weekly
beach water quality info for over 600 locations
along the West Coast, as well as implementing
a daily predictive modeling—or "nowcasting"—
program for five beaches in Southern California.
189
-------
U.S. EPA's 2016 Recreational Waters Conference
CALIFORNIA BEACH
WATER QUALITY
PREDICTIVE
MODELING PROJECT
HEAL THE BAY; STANFORD UNIVERSITY, AND UCLA
aBWjjgjll J. ALAMILLU, M. IAUUARI, A. I HUE, A. BUfcHM, M UULU
Project Outline
Phase 1: Proof of concept
• Can models be developed for CA marine beaches?
Phase II: Pilot at 3 beaches
• Can models be readily integrated in existing M&PN
programs?
Phase 111: Program Implementation
• Developing and Apply ing models lor 20-25 beaches in
Why do we need predictive
models in California?
• Our current monitoring and public notification (M&PN)
programs leave the public at nsk:
* 24-48 hours from sample to posting
.
• Rapid detection methods slill lake hours
lipH
• Of the -500 beaches monitored in CA:
~30 sampled 5x per week
L
~30 2x per week
SB yBasi
>400 only once pci week
Phase I: Proof of Concept
• Completed 2012-2014 at 25 beaches in CA
• 6 years of historical data
• Input factors: rain, tide, wind, solar radiation, etc.
• 5 model types
• 3 FID, 2 seasons
• Summer (S) April to (Jrtoher
• Winter (W) November to March
• Over 700 models developed and tested
• Calibration (2006-2010) and validation (2011-2012'j
%} „
r s
LOS - J^"—• \ I
Angeles"0"" I ***""•
Phase 1: Conclusions
• Models can improve sensitivity while maintaining a reasonable
specificity
• Sensitivity* the ability of a model to accurately predict beach postings
• Specificity: The accurate prediction of open beach days
• Two peer-reviewed scientific, papers were published based on
Phase 1 results
190
-------
Phase TI: Pilot at Three Beaches
• Completed 2015
• Objectives:
• Optimizing models from
Phase I
• Feasibility of using
models w/iu M&PN
programs at CA
beaches
J
Day Three: Session 10
Current Monitoring and Public
Notification at Pilot Beaches
OCHCA
SOCWA
01 Angeles IAHPH-FMD
OCHCA 2/oervveek
Single
Sample or
Ceo Mean
Single
sample
l/Totol and
Fecal;
if tntero
Arroyo Burro
eeacrt
Doheny State
Beach
Santa Monica
Pier Beach
Pilot Design
• Prediction tool: optimized MLR model in
an Excel spreadsheet
• Prediction of post/no-post daily by 10 am Hjffilil
• Study period: Memorial Day to I ,abor Day I
• Tliree Beaches
* LADPH/City of LA Santa Monica Pier
• OCHCA Doheny
SBCEHS Arroyo Burro
Doheny Rear.1i
Pilot Daity Modeling Steps
Obtaining FIB data
• Collect online
environmental data
Run each FIB model
• Cross-check between
with agencies, HtB,
and .Stanford
Posting results online
Posting Predictive Model
Results Online
Heal the Ray Main Webpage
Orange County Website
Posting Predictive Model
Results Online Jk&amsLa
Beach Report Card website:
• Beaehgoers could find wet and dry
grades, Advisories, Noweast results, and
historical information by beach
Both the BRC and OC websites also had FAQ
sections for questions/concerns
Water Quality Noweast:
191
-------
ri
jLa
U.S. EPA's 2016 Recreational Waters Conference
Doheny State Results
Number of samples: 34
Number of predictions: 106
Observed exeeedances: 3
Model captured. 3 vs 2 (current)
False alarms: 1 vs 3 (current)
Phase II Outcomes
• Models successfully run 106 consecutive days
w/ results ready by 10 am
• Daily notification to beach-goers in the morning
everyday including weekends
• 3 agencies voluntarily participated & donated
islaff/i csuuiLCh to the pilot
—• * ;
• Models nan he successfully integrated into
existing M&PN programs
Benefits of Predictive Models
• Improved accuracy in public notification over current method
• Improved understanding of Fl R pollution at the beach and how to
mitigate sources
• Easy and flexible model implementation can be run by the health
agencies or a third party (Heal the Bay)
Phase III: Develop CA
Nowcast System
• SWRCB grant to build permanent CA
Beach Nowcast system:
• Heal the Bay, Stanford, and UCLA
• 3 year roll-out
• 20 summer AB-411 beaches
• winter surf beaches
Phase 111: Develop CA
Nowcast System
• Technical Advisory Committee and
Implementation Advisory Committee
• Public outreach
• On-the-beach public notification program
• Wcbpagc and Mobile Apps
Acknowledgements
Santa Barbara County Environmental Health Services (SBCEHS)
Willie Brummett, Lawerence Fay, and David Brummond
Orange County Health Care Agency (OCHCA)
I arry Rrmnlp"- (lanrpn Harrhjtian Aiwvra. Jnp Qrrman, and Dan Ynknyama)
City of Los Angeles, Environmental Monitoring Division (EMD)
laannice Lcc (Victor Ruiz, Ar-ibe.- Kuhn.Tornmy Nguyen), Mas Dojiri, and Stan Asato
Los Angeles County Department uf Public Health
lacquelinelaylor, Maurice Panto(a,and Nick Brakband (Zepur Ohannessian)
USGS Ohio Water Science Center
Donna Francy
192
-------
193
-------
U.S. EPA's 2016 Recreational Waters Conference
Predictive Modeling and Forecasting of
Water Quality at Recreational Beaches
along Gulf of Mexico Coast
Zhiqiang Deng
Louisiana State University
Abstract
A series of predictive models has been
developed by Louisiana State University for rec-
reational beaches that have experienced frequent
advisories over the past 10 years. The beaches
used in the project, which was funded by the
National Aeronautics and Space Administration
(NASA), were Siesta Key Beach and Venice Beach
in Florida, Orange Street Pier/Park Beach in
Alabama, Harrison County Beach in Mississippi,
Holly Beach in Louisiana, and Galveston Bay
Beach and Corpus Christi Bay Beach in Texas.
The models were constructed using an artificial
neural networks toolbox in the MATLAB pro-
gram and can predict either daily enterococci
levels in beach waters or risks of water quality
standard violations at a beach site as long as
daily data are available for the environmental
parameters (e.g., rainfall, salinity, temperature,
wind, tide [or gage height], and solar radiation).
Some models require less data and some of the
data can be replaced with NASA satellite data.
The models were able to explain 70-86% of the
variations in observed enterococci levels or rec-
reational water quality advisories issued by state
beach monitoring programs. User manuals for
state beach monitoring personnel explain how to
use the models for real-time monitoring of recre-
ational water quality. This presentation will pro-
vide an overview of the models and their perfor-
mance in predicting water quality at the beaches.
It is expected that the adoption and sustained
use of the models will significantly improve the
effectiveness of recreational water programs and
provide better protection of public health in the
Gulf of Mexico states and the nation.
Biosketch
Dr. Zhiqiang Deng is a professor of water
resources engineering at Louisiana State
University. He specializes in predicting and
preventing the contamination of water bodies
with high public health and economic impacts
(primarily recreational beach waters, oyster
harvesting waters, and rivers) through sensor
network-based monitoring, watershed-based
modeling, and sustainability-based mitigation.
Dr. Deng has published over 50 refereed journal
papers in those areas.
194
-------
f*"!
Day Three: Session 10
J
iL2a
Using Probabilities of Enterococci Exceedance
and Logistic Regression to Evaluate Long-Term
Weekly Beach Monitoring Data
Jay Fleisher, PhD
Nova Southeastern University
Abstract
Recreational water quality surveillance
involves comparing bacterial levels to set
threshold values to determine beach closure.
Bacterial levels can be predicted through mod-
els which are traditionally based on multiple
linear regression. The objective of this study
was to evaluate exceedance probabilities—as
opposed to bacterial levels—as an alternate
method to express beach risk. Data were incor-
porated into a logistic regression to identify
environmental parameters most closely cor-
related with exceedance probabilities. The
analysis was based on 7,422 historical sample
data points from the years 2000-2010 for 15
beach sample sites in south Florida. Probability
analyses showed which beaches in the data set
were most susceptible to exceedances. No yearly
trends were observed nor were any relation-
ships to monthly rainfall or hurricanes appar-
ent. Results from logistic regression analyses
found that among the environmental param-
eters evaluated, tide was most closely associ-
ated with exceedances, with exceedances 2.475
times more likely to occur at high tide than at
low tide. The logistic regression methodology
proved useful for predicting future exceedances
at a beach location in terms of probability and
modeling water quality environmental param-
eters with dependence on a binary response.
Beach managers can use this methodology for
allocating resources when sampling more than
one beach.
Biosketch
Dr. Jay Fleisher received his bachelor of
science degree in environmental health science
and master of science degree in environmental
science from the City University of New York,
his master of science degree in epidemiology
from Columbia University's School of Public
Health, and his doctorate in environmental
epidemiology/biostatistics from the Institute of
Environmental Medicine, New York University.
Dr. Fleisher holds faculty positions at Florida's
Nova Southeastern University and University
of Miami. Dr. Fleisher's research interests are in
the fields of chronic and infectious illnesses. He
has focused his research efforts on the health
effects of exposure to waters contaminated with
domestic sewage, indicator organism variabil-
ity, indicator organism-pathogen relationships,
risk assessment, statistical water quality sam-
pling protocols, assessing compliance, setting
of microbial water quality standards, popula-
tion health burden assessment, risk perception,
and risk vs. current standards. Dr. Fleisher
has advised numerous international commit-
tees, organizations, and government agencies
on various aspects of these recreational water
quality issues. In addition, he has authored over
70 peer-reviewed publications and six book
chapters.
195
-------
METHODS
608 samples were utilized in this analysis 10 major environmental
variables and several FIO's were collected on each sample date Both
types of models were run on these data
Multiple Least Squares Linear Regression Vs. Multiple Logistic Regression
A Comparison
ENVIRONMENTAL VARIABLES
> pH
> Salj.rT
> tV'aiK Twcfenr-ra
> T*d*l Saj»
> Twtodxty
> Aaocat of Riaftli in pj*c#daia 5 bocn poor to uucpfcaj
> Axncass of Kirr.fiII is prc>:««iuix I* boon of taapunf
> Wad Dirtcboa
> Wad Sp**d
> Solar R*ii»ncs
"All Environmental Variables entered in Both Models and Backward Selection Procedure used in all models
Results Least Squares Model
Variable Estimate Error SS F Value Pr>F
¦Issicwl -0.21611
¦Tunf«tars 0.0*383
•T>3# 105941
•Raie24HrPrjw -0.02234
¦Wind dasctca
¦Uini Sp«d
0.24457 0.62
11913T1 32^9
2602290 6649
10.83001 27.72
O.M133 0.00031043 3.6102" 14.33
-0.11525 0.01280 31.70014 80.99
036244
0.01320
0.12993
D.0C428
•Sole radiaban -0.00108 O/OOI. S34 2474859 63.23 <.
^Model R ^ftare = QjT^
0.4296
-e.OOOl
<.00C1
<0001
0.0002
-e.0001
.0001
Multiple Linear Least Squares Regression
196
-------
Jej*
Day Three: Session 10
Multiple Logistic Regression
ENVIRONMENTAL VARIABLES
> pK
> Salxin- Waur
> Tl'ctc T«9sparir3r«
> TjdilSttj*
> Tsbiditr
> AaMM of Ri-.r.fiil to th» px*c*ii&g $ kcun pnor to umpLsj
> Aaoat of S-isfi'J is ths 2* bc^n of tasnplas
> Wac Dirfr:tian
> VW Spwd
> Sclir 2t*di*aea
'All Environmental Variables entered in Both Models and Backward Selection Procedure used in all models
Results Logistic Regression Above or
Below Single Sample Criteria
Standard
Wald
Parameter
DF Estimate Error
Chi-Square
Pr > ChiSq
1 -19.1244 4 9439
14.963*
0.0001
• 5*hsitv
1 03526 0.1170
90784
00026
¦ ?ftap«rsrurs
1 0 2816 00663
81041
0.0044
Tad#
1 2 8824 0 7643
14.2163
0.0002
• Solvraduooe
1 -0.00330 D.CCCJ60
380301
-c.0«l
Our Best Multiple Least Squares Regression was computed with a R Square value
of 0.26, while the Multiple Logistic Regression Model yielded a maximum
Sensitivity of 72.9H and a maximum Specificity of 65.9% it a cut point = 0.1- A
backward selection routine was used m both the Logistic and Least Squares
ModeL
197
-------
U.S. EPA's 2016 Recreational Waters Conference
Application of Logistic Regression to
Historical Data
Used Different Data than Previous
• 7,422 Samples analyzed
• Data from 2000-2010
• Data from 13 South Florida Beaches
cfc. TDOHHtaUp S*xlwj P>*ram. ioxA nxm. anJOPS
location of umptinj pom
.i-ta , t. i.t.uMW.-. it « »¦ ta e.i—
///////////
Figure 2. Historical DOH beach sampling from June 2000¦ December 2010 (n = 7.422).
Gray line delineates any samples abo\v the 104 CFL'EPA exceedance level.
Figure 4 Monthly historical exceedance counts
(ban) and their probability (diamonds) for ten
years of exceedance count data
Figures South Florida Water Management
District (SFWMD) Dade County Average ratrfall
from 2000 to 2010 vj. probability and number of
exceedance counts in each month Grey area
represents wet season and white area represents
dry season
Figure 6 Historical exceedance counts (bars) and their probability (diamondsJ by yet
The probabilities were connected to show the \-artabtIity betw een the years
198
-------
• Analysis identified which beaches were most susceptible to exceedances
• Logistic regression proved useful for predicting the probability of an exceedance
• Tide was most closely associated with exceedance
• Results can be used to allocated beach sampling resources
Odds Ratio Estimates
Effect Point Estimate 95% Wald
Confidence Limits
Tidal Conditions lvs3 2.475 1.661 3.687
Tidal Conditions 2 vs3 1.252 0.866 1.811
Table 1. Odds ratio estimates between tidal conditions as computed from Logistic Regression. Tidal
conditions as reported by the FDOH are coded as 1 = High Tide, 2 = Slack Tide, and 3 = Low Tide
199
-------
U.S. EPA's 2016 Recreational Waters Conference
Development of a Predictive Spatial Model to
Understand the Connection between Rainfall
Events and Beach Water Quality
Lance Larson, PhD
Natural Resources Defense Council
Abstract
Throughout coastal portions of the United
States, rainfall events are physical mechanisms
that deliver various urban and rural pollutants
to coastal waterways, threatening human and
ecosystem health. The objective of this research
was to correlate historical beach water quality
exceedances to rainfall events. We developed
a spatial and temporal beach water quality
exceedance model, which queries a database
consisting of water quality sample results
collected over a 10-year period (2005-2014) at
over 8,000 U.S. beaches in 30 states. The model
consists of a series of dynamic database queries
based on a set of user-defined input parameters.
In the database, each water quality sample
record is associated with precipitation totals
recorded on the sample collection date, as well
as for each of the 3 days prior to that sample
date, as measured by the nearest weather sta-
tion submitting data to the National Oceanic
and Atmospheric Administration's Quality
Controlled Local Climatological Data (QCLCD).
Our results suggest a strong connection at the
national, state, county, and beach scale between
increased rainfall events and beach exceedance
occurrences. For example, at the national level,
the failure rate increased from 9% to 21% when
a rainfall event greater than 0.5 inches was
observed within 10 miles within 1 day. Other
states and counties observed disproportionate
changes in exceedance failure rates. Our model
aims to significantly increase our understand-
ing of rainfall influences on beach water quality
throughout the United States, improve water
quality sampling frequencies and planning,
and examine the effectiveness of implementing
watershed pollution reduction strategies.
Biosketch
Dr. Lance Larson is a science center fellow
with the Natural Resources Defense Council in
Washington, DC. He earned a bachelor of sci-
ence degree in environmental engineering from
the California Polytechnic State University in
San Luis Obispo (2008) and a master of science
degree from the South Dakota School of Mines
and Technology (2010). Dr. Larson received a
dual doctorate in environmental engineering
and biogeochemistry from Pennsylvania State
University (2013). His graduate research focused
on acid mine drainage, arsenic and uranium
fate and transport, and biogeochemical inter-
actions between surface and groundwater.
Dr. Larson currently is working with the Land
and Wildlife, Nuclear, and Water programs to
protect U.S. water resources.
200
-------
Day Three: Session 10
Question & Answer Session
Comment 1
(Unknown): For Lance [Larson]. Good database and good work—I am glad you put in the lag times
which are so important in rivers and runoff, and also the saturation of the soil which affects the effect of
rainfall. When you have a 0.9-inch rain event, we consider the storm surge as well as the amount. It's
interesting to take that into account.
Answer 1
Lance Larson: This work raises many more questions than answers. We can use it to build in
other things like that.
Question 2
(Unknown): For Lance [Larson]. How did you make sure the rainfall is in the right area and not in
another watershed?
Answer 2
Lance Larson: You could decrease that distance, so 10 miles would be your threshold. Within
that, it picks it up. It's the threshold cutoff. If we can find the nearest location, we do. We can
run them again at different locations. We did a sensitivity analysis as well.
Answer 2 (follow-up)
Adam Mednick: You said the magic word, "tide," which is very important for incorporating
into models. Also exceedance. In the best practices document we are putting out it's about
probability. One use for VB [Virtual Beach] and modeling is figuring out when whether and
how to test. Glad Jay [Fleisher; made that point during his presentation.
Answer 2 (follow-up)
Mike Cyterski: In terms of Virtual Beach, I'd like to add some other tools, like logistical regres-
sions and neural nets, and lasso regression (where you minimize the number of variables
that you use in your regression).
201
------- |