This document discusses using data mining techniques like artificial neural networks (ANN) to model the impacts of nonpoint source pollution in complex coastal estuaries. ANN models can extract relationships from large monitoring datasets to better understand estuary dynamics and the effects of factors like rainfall, tides, and freshwater flows. Case studies of the Cooper River and Beaufort River estuaries show ANN models accurately simulate dissolved oxygen levels and salinity intrusion in response to these drivers. Data mining allows interactive "what if" scenarios to inform total maximum daily load and permitting decisions.

1. Effects of Nonpoint Source
Marsh Loading on Complex
Estuaries
Edwin A. Roehl, Jr.
John B. Cook, PE
Advanced Data Mining Intl
Greenville, SC

2. South Carolina coastal estuaries
Myrtle Beach
Grand Strand
Georgetown
Charleston
Beaufort
Savannah

3. A brief review of tidal dynamics
Freshwater
Riverine Coastal
Inputs Inputs
Saltwater
Saltwater-Freshwater
Interface
“…estuaries may never really be steady-state
systems; they may be trying to reach a balance
they never achieve.”
Keith Dyer, from Estuaries – A Physical Introduction (1997)

4. Difficult to wrestle down nonpoint
source effects
 Difficult to measure and predict NPS impacts on
upland areas
 Data sets sparse as compared to point source data
 Equations and models to estimate loads can have large
prediction errors (50-100%)
 NPS problem compounded on the coast
 Low-gradient system with little or no slope
 Tidal complexities of receiving stream
 Poorly defined drainage areas
 Limited understanding of runoff process along the coast

5. Complex forces on a tidal river
Overland flow
from watershed
•Small
contributing
watershed
•Little Tidal forcing
freshwater from ocean
inflow
connection
•Tidally
dominated

6. Consider alternative approach to
NPS modeling
 Data mining
 Transforming data into information
 Amalgamation of techniques from various
disciplines: information theory, signal processing,
statistics, machine learning, chaos theory,
advanced visualization
 The physics is manifested in the data
 Need to extract the information from large
data sets of continuous monitoring w/in
estuary

7. Artificial Neural Networks (ANN) models
Mathematical representation of the brain
 provides complicated behaviors from “simple” components
- neurons and synapses
x1
x2 y1
inputs x3 outputs
y2
x4
x5
 models created by training the ANN to learn relationships
between variables in example data
 form of machine learning from Artificial Intelligence (AI)

8. 3D response surfaces for SC, WL, Q
 Surface created
by ANN model
 “Unseen”
variables set to
constant value
 Manifestation of
historical
behavior of
system
 Provides insight
into the process
dynamics or
physics

9. ANN model performance for
hydrodynamic behavior

10. Data mining NPS – Consider Cooper
River Estuary case study
 Sensitivity of DO to rainfall, water
tidal-level flushing action and tidal
range determined
 Model able to simulate rainfall
effects/amounts
 System had long-term data bases
 >3 years of 15-minute WL, DO, SC,
WT

11. Cooper River
Estuary
Area of no
development
Little impact
from all point
sources

12. Signal decomposition of water level
Periodic component
– Tidal range
Chaotic component –
Filtered water level

13. Dissolved oxygen (DO) dynamics
Measured DO time
series
Dissolved-oxygen deficit
= difference b/w saturation
and measured

14. Or, in equation form:
DO deficit (DOD) =
DO [saturated f(T and salinity)]
- DO (measured)

15. Effects of rainfall Z-axis – DOD
on Cooper River X & Y axis – 1- and 3-
day rainfalls
∆2 mg/L
The sensitivity of DOD
to rainfall :
DOD/inch ≈ 2 mg/L/ 8 in.
2 inches
of rainfall over 2 days
2 in ches
= 0.25 mg/L per inch of
rainfall.

16. Cooper River measured and predicted
DO-deficit (DOD) as result of rainfall only
RAINAA=2-day moving window average

17. In addition to rainfall effects, response
surfaces show effects of WLs on DOD
1st response surface shows “Low WL” = higher DOD
(range of 3.0 to 4.5 mg/L)
2nd response surface shows “High WL” = lower DOD
(range of 1.5 to 2.8 mg/L)

18. Data-Driven model’s accuracy, Cooper R.
• Mixing - Tides, Flows from 3 Rivers
• Weather (T, P Dew Point)
• Point Discharge Wastewater
Treatment Plants
• Non-Point Discharges - rainfall,
50% overbank storage
Measured Neural Network BRANCH/BLTM
10 32
9 30
28
Dissolved oxygen (mg/L)
8
Water temperature
(degree Celsius)
26
Temperature
7
24
6
22
5
20
4 18
Dissolved oxygen
3 16
8/21/93 0:30 8/22/93 0:30 8/23/93 0:30 8/24/93 0:30
Date and time

19. Beaufort River
Estuary
Complex tidal system
>9 foot tide range
Net flow to the north
Model developed for
TMDL and NPDES permits
Model simulates 3.5 years
of historical conditions

20. Decision Support Systems make “what-ifs”
easy for Beaufort River TMDL

21. Savannah Harbor deepening
 Model
hydrodynamics
 How far does
salinity intrude
when Harbor is
deepened?
 What happens
when fresh water
flows are low?

22. Accuracy insights: EFDC vs. ANN model
for Savannah River, GA
Salinity, Practical Salinity
Streamflow (cfs)
EFDC R2=0.10
M2M R2=0.90
Units
EFDC unable to
Source: Conrads, P., and Greenfield, J., (2008) predict peaks

23. Simulate reduced freshwater flows with
EFDC and ANN model and compare
EFDC R2=0.10
Salinity, Practical Salinity
M2M R2=0.90
Streamflow (cfs)
Units
Source: Conrads, P., and Greenfield, J., (2008)

24. Summary for NPS Estuary Modeling
 Stormwater and tidal effects (as well as point
source impacts) can be quantified using Data
Mining techniques
 3D visualization gives valuable insight into
process physics of the system
 Data Mining can be used with traditional
approaches to minimize errors in load
estimates from NPS

25. Questions
Contact:
John B. Cook
Advanced Data Mining
Intl; Greenville, SC
John.cook@advdmi.com
843.513.2130
www.advdmi.com