Hypoxia in Narragansett Bay Workshop Oct 2006 “Modeling” In the Narragansett Bay CHRP Project Dan Codiga, Jim Kremer, Mark Brush, Chris Kincaid, Deanna.

Hypoxia in Narragansett BayWorkshop Oct 2006

“Modeling”In the Narragansett Bay

CHRP Project

Dan Codiga, Jim Kremer, Mark Brush, Chris Kincaid, Deanna

Bergondo

Does the word “Model” have meaning?

• Hydrodynamic

• Ecological

• Research vs Applied

• Prognostic vs Diagnostic

• Heuristic, Theoretical, Conceptual, Empirical, Statistical, Probabilistic, Numerical, Analytic

• Idealized/Process-Oriented vs Realistic

• Kinematic vs Dynamic

• Forecast vs hindcast

CHRP Program Goals (selected excerpts from RFP)

• Predictive/modeling tools for decision makers

• Models that predict susceptibility to hypoxia

• Better understanding and parameterizations

• Transferability of results across systems

• Data to calibrate and verify models

Following two presentations

Our approaches• Hybrid Ecological-Hydrodynamic Modeling

– Ecological model: simple• Few processes, few parameters• Parameters that can be constrained by measurements• Few spatial domains (~20), as appropriate to measurements available• Net exchanges between spatial domains: from hydrodynamic model

– Hydrodynamic model: full physics and forcing of ROMS• realistic configuration; forced by observed winds, rivers, tides, surface fluxes• Applied across entire Bay, and beyond, at high resolution• Passive tracers used to determine net exchanges between larger domains of

ecological model

• Empirical-Statistical Modeling– Input-output relations, emphasis on empirical fit more than mechanisms– Development of indices for stratification, hypoxia susceptibility– Learn from hindcasts, ultimately apply toward forecasting

Heuristic models in research: iterative failure = learning

ConceptualModel

Runs that fall short

Processes

Formulations

Parameter values

But for management models:• Heuristic goal less impt

• Accurate even if not precise

• Well constrained coefs• Simple (?) (at least understandable)

_____________________________ ≠ Research models

A paradox --

“Realism” = many parameters weakly constrained limited data to corroborate

i.e. “Over-parameterized” (many ways to get similar results)

:. Accuracy is

unknown. (often unknowable)

An alternative approach? 4 state variables, 5 processes

N P

N

Land-use

Atmosphericdeposition

N P

Productivity

Temp, Light,Boundary Conditions

Chl, N, P, Salinity

Phytoplankton

Sedimentorganics

.

.

Physics

Surface layer

- - - - - - - - -

Deep layer

- - - - - - - - -

Bottom sediment

O2

Flux tobottom

Photic zoneheterotrophy

Benthicheterotrophy Denitri-

fication

O2 coupledstoichiometrically

Processes of the model (excluding macroalgae...)

mixing flushing

Corroboration:“Strength in numbers”

Shallow test sites (MA, RI,

CT)

Long Island Sound -- Hypoxia

August 20

Deep test sites (MA, RI, CT, VA, MD)

Chesapeake Bay

Narragansett Bay

Long Island Sound

Initial Conditions

Forcing Conditions

Output

EquationsMomentum balance x & y directions:u + vu – fv = + Fu + Du t xv + vv + fu = + Fv + Dv t yPotential temperature and salinity :T + vT = FT + DT

t S + v S = FS + DS

t The equation of state:= (T, S, P) Vertical momentum: = - gz o

Continuity equation:u + v + w = 0x y z

Hydrodynamic Model

ROMS Model

Regional Ocean Modeling System

Grid Resolution: 100 mGrid Size: 1024 x 512Vertical Layers: 20River Flow: USGSWinds: NCDCTidal Forcing: ADCIRC

Open Boundary

Hydrodynamic Model

This project: Mid-Bay focus

Extent of counterMt. Hope Bay circulation/exchange/mixing study. ADCP, tide gauges (Deleo, 2001)

Bay-RIS exchange study (98-02)

Narragansett Bay Commission: Providence & Seekonk Rivers

Summer, 07: 4 month deployment (Outflow pathways)

This project: Mid-Bay focus

Extent of counterMt. Hope Bay circulation/exchange/mixing study. ADCP, tide gauges (Deleo, 2001)

Bay-RIS exchange study (98-02)

Narragansett Bay Commission: Providence & Seekonk Rivers

Summer, 08: Deep return flow processes

Model-Data Comparison

Salinity - Phillipsdale

Sal

inity

(pp

t)

Time (days)

Model

Data

Model-Data ComparisonShallows: North-South Component

-0.25

-0.15

-0.05

0.05

0.15

10 15 20

Time

Model Velocity

(m/s)

-250

-150

-50

50

150

Observed Velocity

(mm/s)

Bottom-model

Bottom

Channel: North-South Component

-0.25

-0.15

-0.05

0.05

0.15

0.25

10 15 20

Time (days)

Model Velocity

(m/s)

-250

-150

-50

50

150

250

Obsevered Velocity (m/s)

Bottom-model

Bottom

dP1/dt = P1(G-R) - k1,2P1V1 + k2,1P2V2 ...

Hybrid: Driving Ecological model with Hydrodynamic Model:

Lookup Table of Daily Exchanges (k)

DYE_08

DYE_02DYE_03

DYE 05

DYE_01

DYE_09DYE_07

DYE_06 DYE 04

Modeling Exchange Between Ecological Model Domains

Passive Tracer Experiment

Passive Tracer Experiment

Passive Tracer Experiment

Long-term Aims:Hybrid Ecological-Physical Model

• Increased spatial resolution of ecology: approach TMDL applicability

• Scenario evaluation– Nutrient load changes– Climatic changes

• Alternative to mechanistic coupled hydrodynamic/ecological modeling

Empirical/Statistical ModelingOverall Goals

• Data-oriented—complements Hybrid– less mechanistic• Synthesize DO variability

– Spatial (Large-scale CTD; towed body)

– Temporal (Fixed-site buoys)

• Develop indices– Stratification

– Hypoxia vulnerability

• First: Hindcasts to understand relationship between forcing (physical and biological) and DO responses

• Long-term: Predictive capability for forecasting and scenario evaluation

• Candidate predictors for DO– Biological

• Chlorophyll• Temperature & solar input• Nutrient inputs (Rivers, WWTF, Estuarine

exchange) • Others

– Physical • River runoff, WWTF water transports• Tidal range cubed (energy available for mixing)• Windspeed cubed (energy available for mixing)• Others (Wind direction; Precip; Surface heat flux)

Strategy: start simple & develop method

• Start with Bullock Reach timeseries– 5 yrs at fixed single point (no spatial information)

• Investigate stratification (not DO-- yet)– Target variable: strat = [sigt(deep) – sigt(shallow)]– Include 3 candidate predictor variables:

• River runoff (sum over 5 rivers)

• Tidal range cubed (energy available for mixing)

• Windspeed cubed (energy available for mixing)

2001

Visually apparent features

• Stratification reacts to ‘events’ in each of:– River inputs– Winds– Tidal stage

• Stratification ‘events’ appear to be– Triggered irregularly by each process– Lagged by varying amounts from each process

Low-pass and subsample to 12 hrs…Compare techniques

• Multiple Linear Regression (MLR)– No lags– Optimal lags – determined individually

• Static Neural Network– No lags– Lags from MLR analysis

• [coming soon] Dynamic Neural Network– Varying lags– Multiple interacting inputs

Multiple Linear Regression

No lagsr2=0.42 (River alone: 0.36)

Observed Model

MLR with lags River 2 days Wind 1 day Tide 3.5 days

r2=0.51 (River alone: 0.48)

Stratificationt [kg m-3]

Static Neural NetNo lags

r2=0.55 (River alone: 0.41)

Static Neural NetLags from MLR

r2=0.59 (River alone: 0.52)

Advantages/Disadvantagesof Neural Networks

• Advantages– Nonlinear, can achieve better accuracy– Excels with multiple interacting predictors; – Dynamic NN: input delays capture lags

• Varying lags from multiple interacting inputs

– Transferable; conveniently applied to other/new data– Easy to use (surprise!!)

• Main disadvantage– opaque “black-box” can be difficult to interpret;

ameliorated by: complementary linear analysis, sensitivity studies, isolating/combining predictors

Next steps

• Stratification– Consider additional predictors:

• Surface heat flux; precipitation; WWTF volume flux

– Different sites (North Prudence, etc)– Treat spatially-averaged regions

• Apply similar approach to DO– Finish gathering forcing function data

• Chl; solar inputs; WWTF nutrients

– Corroborate Hybrid Ecological-Hydrodynamic Model