Title: Predictive Skill, Predictive Capability and Predictability in Ocean Forecasting
1Predictive Skill, Predictive Capability and
Predictability in Ocean Forecasting
- Allan R. Robinson
- Patrick J. Haley, Jr.
- Pierre F.J. Lermusiaux
- Wayne G. Leslie
- Division of Engineering and Applied Sciences
- Department of Earth and Planetary Sciences
- Harvard University
- 29 October 2002
2Ocean Prediction
- Ocean prediction for science and operational
applications has been initiated on basin and
regional scales. Evaluation is now essential. - Predictability limit is the time for two slightly
different ocean states to evolve into realistic
but entirely different states - Predictive capability is ultimately limited by
predictability but errors in data, models and
methodology now limit prediction capability to
shorter times
3Predictive Skill
- Qualitative and quantitative evaluation of ocean
forecasts by generic and regional-specific skill
metrics is essential - Phase errors/structural errors, initial/BC/model
errors and their sources need to be identified - Simple metrics from meteorology (root-mean square
error, anomaly pattern correlation coefficient)
used now but more sophisticated statistical
metrics and quantitative measures associated with
underlying dynamical processes are required.
4(No Transcript)
5CPSE/REA
- Coastal Predictive Skill Experiment (CPSE)
- measures the ability of a forecast system to
combine model results and observations in coastal
domains or regimes and to accurately define the
present state and predict the future state - oversampling is required for rigorous
quantitative verification - provides the basis for optimal, efficient
sampling for required accuracies - Rapid Environmental Assessment (REA)
- defined in the NATO naval environment as "the
acquisition, compilation and release of
tactically relevant environmental information in
a tactically relevant time frame"
6ASCOT-01
- Assessment of Skill for Coastal Ocean Transients
- 6-26 June 2001
- Massachusetts Bay and Gulf of Maine
- http//www.deas.harvard.edu/leslie/ASCOT01/sci_pl
an.html - Predictive Skill Experiment
- quantitative skill evaluation
- forecast system development
- real-time at-sea forecasting
- real-time adaptive sampling
- Coupled Physical-Biological Experiment
- initialization surveys of Mass. Bay and Gulf of
Maine - wind-induced events, e.g. upwelling and buoyancy
circulations - Gulf of Maine inflow to Mass. Bay
- Mass. Bay outflow to Gulf of Maine
- MWRA diffuser dispersion
- verification survey of Mass. May
- Multiple vessels
- NRV Alliance SACLANTCEN La Spezia,Italy
- RV Gulf Challenger UNH Portsmouth, NH
7ASCOT-01 Data and Modeling Domains 6-26 June 2001
8Forecast Skill Metrics Skill of the operational
forecasts is measured using the metrics,
Root-Mean-Square Error (RMSE) and Pattern
Correlation Coefficient (PCC). These numbers are
computed model level by model level (1 to 16),
and as a volume average. Perfect values of the
RMSE and PCC are, respectively, zero and one.
The metrics RSME and PCC are respectively
defined by where denotes the true
ocean, its forecast, a background field
vector (e.g. large-scale field, climatological
field, etc.), and . 2 the vector l2
norm. A classic measure of skill is to compare
the RMS and PCC of the forecast with that of the
initial conditions (IC) (persistence). If the
RMSE of the forecast is smaller than that of the
IC, the forecast has RMS-skill or beats
persistence. Similarly, if the PCC of the
forecast is larger than that of the IC, the
forecast has PCC-skill or has better patterns
than persistence.
9Observation Errors ASCOT-01 19 June 2001
10ASCOT-01 Skill Metrics RMS (Temperature -
Left Salinity - Right) PCC (Temperature -
Left Salinity - Right)
11ASCOT-02
- Assessment of Skill for Coastal Ocean Transients
- 7-17 May 2002
- Tyhrrenian Sea, Ligurian Sea, Corsican Channel,
Elba - http//www.deas.harvard.edu/leslie/ASCOT02/sci_pl
an.html - Predictive Skill Experiment
- quantitative skill evaluation
- forecast system development
- real-time at-sea forecasting
- real-time adaptive sampling
- rigorous test of distributed ocean prediction
system - AUV exercise support
- Physics Experiment
- initialization survey of Corsican Channel and
Elba island region - flow between Corsica and Elba
- anticyclone north of Elba
- flow between Elba and the coast of Italy
- reduce multi-variate forecast errors
- Multiple vessels
- NRV Alliance SACLANTCEN La Spezia,Italy
12ASCOT-02 Data and Modeling Domains 7-17 May 2002
13Observation Errors ASCOT-02
15 May
16 May
14ASCOT-02 Skill Metrics RMS (Temperature -
Left Salinity - Right) PCC (Temperature -
Left Salinity - Right)
15General AdaptiveSampling Objectives
- Go to dynamical hotspots
- Reduce error variance
- Reduce errors for tomorrow
- Maintain accurate forecast
- Maintain accurate synoptic picture
- Optimal sampling issues
- Automate all 3 above quantitatively
- Nonlinear and interdisciplinary impacts on the
sampling - Optimal sampling can be highly dependent on
objectives and metrics - Reducing error in analysis differs from reducing
error in forecast - Minimal final time error differs from minimal
time-averaged error - Minimize cost function containing 3 terms based
on - forecasted model errors (ESSE),
- forecasted significant dynamical events (MS-EVA,
pattern recognition) - maximum length of time an area can be left
without updating
16Motivations for Adaptive Sampling Tracks
ASCOT-02/GOATS
- Sample in regions not yet covered to locate local
structures - Sample in regions not recently covered to
understand evolution of structures - Determine strength and structure of anticyclone
north of Elba - Determine general nature of flow in vicinity of
Procchio Bay (e.g.. is it from north or result of
flow through Corsican Channel from Tyrrhenian
turning around island?) - Evaluate structure and evolution of flow between
Corsica and Elba - Determine impact of flow between Elba and coast
of Italy
17Adaptive sampling tracks designed on a real-time
basis.
AUV - Procchio Bay
NRV Alliance - Channel Domain
18(Top left) Surface temperature after 4 days of
model run. Overlaid on the temperature field are
the 50, 200 and 500m isobaths. (Right) Satellite
sea surface temperature. (Bottom Left) Surface
current from the ocean model. All fields are
from 3 October.
19Currents measured by NRV Alliance with the
ship-borne ADCP during the first update surveys.
Data of 4 consecutive nights are merged. The SE
current in the south-eastern corner is due to
high winds.
Currents measured by NRV Alliance with the
ship-borne ADCP during update surveys in early
October. The anti-cyclonic eddy has shifted
towards the north.
20HOPS ASCOT-01 Simulations (5 meters), SeaWiFS
Composite Imagery and in situ data
13 June
10-17 June
20 June
18-25 June
21Conclusions
- It is critically important to interpret and
evaluate regional forecasts in order to establish
usefulness to scientific and applied communities - Results from ASCOT highlight the dual use of data
for assimilation and skill evaluation and
demonstrate quantitative forecast skill - Real-time forecast experiments can lead to
discoveries of regional features - Multi-scale adaptive sampling is a fundamental
component of forecast systems
22Issues in MultiscaleAdaptive Sampling
- Uniformly sampled observations for initialization
and assimilation as forecasts advance in time - sampled uniformly over a predetermined
space-time grid, adequate to resolve scales of
interest - only a small subset of observations have
significant impact on the accuracy of the
forecasts - impact subset is related to intermittent
energetic synoptic dynamical events - Adaptively sampled observations for
initialization and assimilation as forecasts
advance in time - sampling scheme tailored to the ocean state to be
observed - knowledge of ocean state from ongoing
observations, nowcasts and forecasts - adaptive sampling targets observations of
greatest impact - efficient adaptive sampling reduce observational
requirements by one or two orders of magnitude - Subjective adaptive sampling and objective
adaptive sampling - sampling can be based on environmental forecasts
or error forecasts - forecast information combined with a priori
experience to intuitively choose future sampling - objectively, forecast serves as input to a
quantitative sampling criterion whose
optimization predicts the adapted sampling - automated objective adaptive sampling