Lyon Lanerolle 1,2 , Aaron J. Bever 3 and Marjorie A. M Friedrichs 4

Office of Coast Survey / CSDL

Sensitivity Analysis of Temperature and Salinity from a Suite of Numerical Ocean Models for the

Chesapeake BayLyon Lanerolle1,2, Aaron J. Bever3 and Marjorie A. M Friedrichs4

1NOAA/NOS/OCS/Coast Survey Development Laboratory,1315 East-West Highway, Silver Spring, MD; 2Earth Resources Technology (ERT) Inc.,6100 Frost Place, Suite A, Laurel, MD; 3Delta Modeling Associates, Inc., San

Francisco, CA ; 4Virginia Institute of Marine Science, The College of William & Mary, Gloucester Point, VA.

U.S. IOOS Coastal Ocean Modeling Testbed

24 January 201110th Symposium on the Coastal Environment

92nd Annual American Meteorological Society Meeting


Introduction and Motivation• Physical component of Numerical Ocean models generate water

elevations, currents, T and S

• Water quality models and ecological models/applications rely primarily on T and S (from the physical model)

• Expect “best” water quality predictions to result from the “best” T and S predictions (relative to observations)

• Therefore attempt to:

examine predicted T, S sensitivity to various model parameters optimize the predictions for T, S from models examine how different models compare with observations and each other employ “best” T, S predictions for water quality forecasting


US IOOS Coastal Ocean Modeling Testbed

• Focus on Estuarine Dynamics and Modeling component

• Ideal candidate is Chesapeake Bay: Extensive data sets available (in time and space) Several numerical ocean model applications available

• Ocean models available for Testbed: CBOFS (NOAA/NOS/CSDL-CO-OPS, Lyon Lanerolle et al.) ChesROMS (U-Md/UMCES, Wen Long et al.) UMCES ROMS (U-Md/UMCES, Ming Li, Yun Li) CH3D (CBP, Ping Wang; USACE, Carl Cerco) EFDC (William & Mary/VIMS, Jian Shen and Harry Wang)

• Observed data – Chesapeake Bay Program (CBP)

• Simulation period – 2004 calendar year (2005 is similar)


Model-Observation Comparison Metrics

• Metric used is the Normalized Target Diagram (Jolliff et al. 2009)

• m’ = m - M, o’ = o - O

• σo - SD of obs.

• Model skill is distance from origin (origin = perfect model-obs. fit)• Graphical versus numerical approach more informative

unbiased RMSD[sign(σm - σo)· {Σ (m’-o’)2 / N}1/2] / σo

Bias [(M-O) / σo]

+1

+1

-1

-1

Overestimates RMSD

Overestimates mean


Chesapeake Bay Program Comparison Stations

• Model(s)-Observation comparisons were made at 28 CBP stations• Stations covered lower, mid, upper Bay, Bay axis and tributaries


Model Calibration / Parameter Sensitivity(using CBOFS)

Bottom T Bottom S

Maximum S stratification Depth of max. S strat.

Greatestsensitivity


Global Errors

Kachemak Bay

Upper Cook Inlet

Nests

Bottom T Bottom S

• Errors were computed by considering all (28) stations at all depths and for full year

• T - CBOFS best with accurate mean and error is in overestimated RMSD• – EFDC and ChesROMS underestimate RMSD and latter underestimates mean

• S – EFDC, CH3D best but have opposite RMSD error; former underestimates mean• - again, errors show greater spread and larger magnitude than for T


Geographical Error Dependence (T)

• Bay axis errors plotted as a function of station latitude

• Errors are for bottom T

• No strong dependence on geography (lower-, mid-, upper-bay) – small error spread

• Different models have different skill characteristics (over/under estimation of mean and RMSD)


Geographical Error Dependence (S)

• Errors are for bottom S

• Unlike T, errors show greater spread

• 3 ROMS models similar, have largest errors and greatest in upper Bay

• CH3D, EFEC smaller errors, evenly spread and less geographical dependence


Value-Based Error Dependence (T)

• Errors for bottom T plotted as the observed mean value itself

• Models show similar trends with UMCES ROMS and CBOFS showing slight improvements over others

• Generally, warmer T values have smaller errors – as seen by UMCES ROMS


Value-Based Error Dependence (S)• Bottom S errors show

greater spread than for T

• Error characteristics from models are similar except UMCES ROMS – full underestimation of mean

• No consistent value-based error dependence in any of the models


Seasonal Error Dependence (T)

• Errors for bottom T plotted as a function of month in 2004

• Spread in errors seen for all models – EFDC the most; warmer months have smaller errors

• CBOFS is most accurate and errors well balanced

• CH3D – overestimates mean, ChesROMS – underestimates mean

• EFDC – largest errors during latter half of year


Seasonal Error Dependence (S)

• Bottom S errors show less spread than for T

• Different error characteristics in each model

• 3 ROMS models show similarity – overestimation of RMSD and underestimation of mean (except CBOFS)

• CH3D – underestimates RMSD

• EFDC – underestimates mean and under- and overestimates RMSD


Conclusions• Inferences for 2004, 2005 similar - so focused on 2004• Bottom S was the most sensitive variable and was used as a proxy• Model calibration/sensitivity study showed CBOFS was not significantly

sensitive to parameter variation• Global T, S errors – no drastic differences between different model

predictions (although some were relatively better)• Geographical error dependence – ROMS models had largest errors in

upper Bay; CH3D, EFDC less geographically dependent• Value-based error dependence – warmer T values have smaller errors;

no discernible error trends for S• Seasonal error dependence – T from ROMS models are similar and

CBOFS has best error balance (mean/RMSD); for S, models show different error characteristics with under/over estimation of mean/RMSD in each

• Target Diagrams proved to be an invaluable and straightforward metric for studying T and S model-observation differences

Lyon Lanerolle 1,2 , Aaron J. Bever 3 and Marjorie A. M Friedrichs 4

Documents