Download presentation
Presentation is loading. Please wait.
1
Experimental System for Predicting Shelf-Slope Optics (ESPreSSO): Assimilating ocean color data using an iterative ensemble smoother: skill assessment for a suite of dynamical and error models Dennis McGillicuddy Keston Smith
2
Ocean color assimilation as a solution to the cloudiness problem State estimation (compositing) Process studies SW06 data – Gawarkiewicz et al.
3
Methodology Forward model (2-D): c: surface layer (20m) chlorophyll concentration v: velocity provided by hydrodynamic model D: diffusivity, 10 m s -2 S(x,y), R(x,y): unknown source/sink terms Data (d): H: linear measurement operator η: measurement error Approach: Utilize an ensemble (N=400) of Monte Carlo simulations to estimate initial conditions c(t 0 ) and source/sink terms S(x,y), R(x,y)
4
Methodology cont’d Best prior estimate: Initialize with (climatology) Source/sink term S(x,y)=0; R(x,y)=0 Posterior estimates of the ICs and source sink terms are Gaussian perturbations about the best prior estimates of ci(t0), S(x,y), and R(x,y) lead to an ensemble of simulations Kalman gain computed from Monte Carlo approximations to the covariance between the unknown parameters and the model prediction of the observations
5
Use of an explicit biological model Phytoplankton: Nutrients: Satellite ocean color data samples the phytoplankton field:
6
Model domains Shelf-scale ROMS (9km resolution) He and Chen, submitted Assimilation subdomain Region of interest
7
Mean of the prior initial conditions: MODIS climatology for August Chlorophyll a – mg m -3
8
Velocity Field 9km ROMS time-mean for Jul 25-Sept 9
9
Results Prior RHS=0 RHS=S(x,y) RHS=R(x,y)c N-P model Satellite data Ernesto
10
Inferred biological fields RHS=S(x,y) RHS=R(x,y)c N-P model Uptake Mortality
11
Skill assessment: sensitivity to observational error Observational error standard deviation RMS misfit to unassimilated data ADR 43% ADS 36% NP 32% AD 18%
12
Comparisons with Gawarkiewicz in situ data
13
Year Day Correlation Model vs. in situ Model vs. satellite Satellite vs. in situ Ernesto
14
Conclusions Ensemble smoothing methodology shows promise Goodness of fit depends on data (amount, underlying phenomenology) parameters of the assimilation procedure λ x, λ obs, σ obs biological dynamics of the forward model Future directions apply to additional ESPreSSO / BIOSPACE field foci in situ data higher resolution time-dependent velocity fields methodological development more sophisticated bio-optical models joint uncertainties in physics and biology skill assessment
15
Extras
16
Sensitivity to observational error: misfit Observational error standard deviation RMS misfit to active data Prior misfit σ obs =0.1
17
Necessity of an iterative approach
18
Methodology (3) Posterior estimates of the ICs and source sink terms are Kalman gain computed from observational error covariance W and the ensemble covariances P, P c, and P R with
19
Methodology (4) Model prediction covariance (prior) Monte Carlo approximations to the covariance between the unknown parameters and the model prediction of the observations: N obs X N obs N unk X N obs
20
Methodology cont’d Best prior estimate: Initialize with (climatology) Source/sink term S(x,y)=0; R(x,y)=0 Gaussian perturbations about the best prior estimates of c i (t 0 ), S(x,y), and R(x,y) lead to an ensemble of simulations Posterior estimates of the ICs and source sink terms are Kalman gain computed from Monte Carlo approximations to the covariance between the unknown parameters and the model prediction of the observations
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.