BioOptical Models and assimilation strategies

1
Bio-Optical Models and assimilation strategies

• Dalhousie group
• Katja Fennel
• Moritz Lehmann
• Paul Mattern
• Susanne Craig
• Michael Dowd

2
Main objective

• Develop and contrast biological and bio-optical
  models and data assimilation strategies that
  improve near-shore predictive capabilities by
  taking advantage of emerging autonomous data.
• Implement and contrast alternative approaches and
  assess their predictive skill.

3
(No Transcript)
4
Lima-Doney model in MABGOM
5
Lima-Doney model in MABGOM
POC
MAB shelf
MAB slope
POC
GS
6
Lima-Doney model in MABGOM
7
Lima-Doney model in MABGOM
8
Lima-Doney model in MABGOM
Deriving IOPs from biomass-based model using
approach of Fujii, Boss Chai (2007)
9
IOP-based model
NO3 SmS O2
DIN

• aSW aNAP
• a

aCDOM
aphy

• bSW bbg
• b

bphy
bdet

• variables/combinations of variables
• are directly observable
• potential for spectral resolution
• potentially improved underwater light field

? aNAP cP
10
IOP-based model
11
Ensemble assimilation

• Ensembles are an approach for dealing with
  uncertainty in numerical prediction
• Models of fluid flow coupled with biological
  processes are highly non-linear
• Reality fundamental limits on accuracy/predictabi
  lity
• Model ensembles approximate true state of the
  system (PDF) by an ensemble which samples
  uncertain inputs and processes predictions in
  form of PDF (probability of different outcomes)

12
Ensemble assimilation

• Model state X
• Ensemble X(i)i1n (n ensemble members)
• Observations y1t (y1,y2, , yt), t1,,T
• Transition from t-1 to t

• Prediction step X(i)t-1t-1 ? X(i)tt-1
• Update step target X(i)tt (different
  methods EnKF, SIR)
• EnKF X(i)tt X(i)tt-1 Kt(Y(i)t -
  HtX(i)tt-1) ?i
• Model-data discrepancy is added to the
  model state weighted by the Kalman gain matrix.
  (Evensen 2003, 2006)
• SIR resampling of forecast ensemble
• Probability is assigned to each ensemble
  member based on its agreement with new
  observation ensemble is resampled given these
  probabilities.
• Hence, ensemble member close to obs. (high
  weight) are likely to be picked, ensemble member
  far from obs. (low weight) is likely to drop out.
  (Ristic et al. 2004)

13
Ensemble assimilation applied to 1-D model at
Bermuda
T (data) T (model)
Mattern et al. (submitted to JMS)
S (data) S (model)
14
Ensemble assimilation applied to 1-D model at
Bermuda
15
Ensemble assimilation applied to 1-D model at
Bermuda
16
Ensemble assimilation applied to 1-D model at
Bermuda
17
Ensemble assimilation applied to 1-D model at
Bermuda
PON
DIN
Chl
18
Ensemble assimilation applied to aCDOM in ESPRESSO
Assimilation period April 28, 2006 to May 25,
2008
aCDOM ensemble mean
aCDOM observation
aCDOM model - obs
19
Ensemble assimilation applied to aCDOM in ESPRESSO
20
Ensemble assimilation applied to aCDOM in ESPRESSO
Evolution of CDOM degradation rate
21
Our next steps

• Nest ESPRESSO within MABGOM (improve IOP b.c.s)
• Refine calculation of likelihood during update
  step (spatial weighting)
• Merge ensemble assimilation with the optimal
  4DVAR physics produced by RU team and assess
  predictive skill
• Include all IOP variables in ensembles
  assimilation, optimize its parameters, and
  transfer IOP model to RU for inclusion into 4DVAR
• Extend assimilation scheme to include glider
  observations
• Adaptive sampling Implement scheme to optimize
  glider paths

22
Thank you!