Ensemble Kalman Filter Geir Evensen and Laurent Bertino Hydro Research Centre, Bergen, Norway, Nansen Environmental and Remote Sensing Center, Bergen, Norway
The Ensemble Kalman Filter (EnKF) Represents error statistics using an ensemble of model states. Evolves error statistics by ensemble integrations. Variance minimizing analysis scheme operating on the ensemble. Monte Carlo, low rank, error subspace method. Converges to the Kalman Filter with increasing ensemble size. Fully nonlinear error evolution, contrary to EKF. Assumption of Gaussian statistics in analysis scheme.
The error covariance matrix Define ensemble covariances around the ensemble mean C f ψψ (Cf ψψ )e = (ψ f ψ f )(ψ f ψ f ) T C a ψψ (Ca ψψ )e = (ψ a ψ a )(ψ a ψ a ) T The ensemble mean ψ is the best-guess. The ensemble spread defines the error variance. The covariance is determined by the smoothness of the ensemble members. A covariance matrix can be represented by an ensemble of model states (not unique).
Dynamical evolution of error statistics Each ensemble member evolves according to the model dynamics which is expressed by a stochastic differential equation dψ = g(ψ)dt + h(ψ)dq. The probability density f then evolve according to Kolmogorov s equation f t + i (g i f) ψ i = 1 2 i,j 2 f(hc qq h T ) ij ψ i ψ j. This is the fundamental equation for evolution of error statistics and can be solved using Monte Carlo methods.
Analysis scheme (1) Given an ensemble of model forcasts, ψ f j, defining forecast error covariance C f ψψ (Cf ψψ )e = (ψ f ψ f )(ψ f ψ f ) T. Create an ensemble of observations d j = d + ɛ j, with d, the real observations, ɛ j, a vector of observation noise, ɛɛ T = (C ɛɛ ) e C ɛɛ.
Analysis scheme (2) Update each ensemble member according to ψ a j = ψ f j+(c f ψψ )e M T( M(C f ψψ )e M T +C ɛɛ ) 1 (d j Mψ f j Thus, the update of the mean becomes ψ a = ψ f +(C f ψψ )e M T( M(C f ψψ )e M T +C ɛɛ ) 1 (d Mψ f). The posterior error covariance becomes (C a ψψ )e = (C f ψψ )e (C f ψψ )e M T ( M(C f ψψ )e M T + C ɛɛ ) 1 M(C f ψψ ) e. ).
Example: Lorenz model Application with the chaotic Lorenz model Illustrates properties with higly nonlinear dynamical models. From Evensen (1997), MWR.
EnKF solution 20 EnKF: Estimate 15 10 5 0-5 -10-15 -20 0 5 10 15 20 25 30 35 40
EnKF error variance 30 EnKF: Error variance 25 20 15 10 5 0 0 5 10 15 20 25 30 35 40
Summary: Lorenz model The EnKF works well with highly nonlinear dynamical models. There is no linearization in the evolution of error statistics. Methods using tangent linear or adjoint operators have problems with the Lorenz equations: limited by the predictability time, limited by the validity time of tangent linear operator. Can we expect the same to be true for more complex models with large state spaces?
Analysis equation (1) Define the ensemble matrix A = (ψ 1, ψ 2,..., ψ N ) R n N. The ensemble mean is (defining 1 N R N N 1/N) A = A1 N. The ensemble perturbations become A = A A = A(I 1 N ). The ensemble covariance matrix (C ψψ ) e R n n becomes (C ψψ ) e = A (A ) T N 1.
Analysis equation (2) Given a vector of measurements d R m, define stored in d j = d + ɛ j, j = 1,..., N, D = (d 1, d 2,..., d N ) R m N. The ensemble perturbations are stored in E = (ɛ 1, ɛ 2,..., ɛ N ) R m N, thus, the measurement error covariance matrix becomes C e ɛɛ = EET N 1.
Analysis equation (3) The analysis equation can now be written A a = A+(C ψψ ) e M T( M(C ψψ ) e M T +C ɛɛ ) 1 (D MA ). Defining the innovations D = D MA and using previous definitions: A a = A + A (MA ) T ( (MA )(MA ) T + C ɛɛ ) 1 D. i.e., analysis expressed entirely in terms of the ensemble
Analysis equation (4) Define S = MA and C = SS T + C ɛɛ. Use A = A(I 1 N ). Use 1 N S T 0. ( ) 1 A a = A + A S T SS T + C ɛɛ D = A + A(I 1 N )S T C 1 D = A (I + S T C 1 D ) = AX
Remarks (C ψψ ) e never computed but indirectly used to determine M(C ψψ ) e M T = SS T. Covariances only needed between observed variables at measurement locations. Analysis may be interpreted as: combination of forecast ensemble members, or, forecast plus combination of covariance functions. Accuracy of analysis is determined by: the accuracy of X, the properties of the ensemble error space.
Remarks For a linear model, any choice of X will result in an analysis which is also a solution of the model.
Examples of ensemble statistics Taken from Haugen and Evensen (2002), Ocean Dynamics. OGCM (MICOM) for the Indian Ocean. Assimilation of SST and SLA data.
Spatial correlations Latitude 25 20 15 10 5 0-5 SST-DP(1) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1-0.1-0.2-0.3-0.4-0.5-0.6-0.7-0.8-0.9-1 Latitude 25 20 15 10 5 0-5 SSH-DP(1) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1-0.1-0.2-0.3-0.4-0.5-0.6-0.7-0.8-0.9-1 -10-10 -15-15 -20 40 60 80 100 120 Longitude -20 40 60 80 100 120 Longitude Latitude 25 20 15 10 5 0-5 SST-SSH 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1-0.1-0.2-0.3-0.4-0.5-0.6-0.7-0.8-0.9-1 Latitude 25 20 15 10 5 0-5 SSH-UVEL 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1-0.1-0.2-0.3-0.4-0.5-0.6-0.7-0.8-0.9-1 -10-10 -15-15 -20 40 60 80 100 120 Longitude -20 40 60 80 100 120 Longitude
Correlation functions SSH SSH SSH SST SSH DP SST SSH SST SST SST DP
Correlation functions SSH SSH SSH SST SSH DP SST SSH SST SST SST DP
Time Depth: Temperature 0 100 200 Depth 300 400 OBSERVED AND SIMULATED TEMPERATURE 32 31.5 31 30.5 30 29.5 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 500 100 120 140 160 180 200 220 240 260 Days