Ensemble data assimilation methods
for merging remotely sensed terrestrial water storage changes with hydrological models
Maike Schumacher, Jürgen Kusche
Institute of Geodesy and Geoinformation, University of Bonn, Germany
Workshop on Data Assimilation in Terrestrial Systems
Bonn, September 19-21, 2016
Canopy Snow
Soil/Vegetation Surface
Water River
Groundwater
Simplification of Reality
2
+ Precipitation - Evapotranspiration
- Runoff
Terrestrial Water Cycle
Large-Scale Hydrological Modelling
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
Limitation of models - parameters
- forcing data (precipitation)
- model structure
3
Canopy Snow
Surface Water
River
Groundwater
Total Water Storage Anomaly (TWSA)
Altimetry
Groundwater well Soil moisture
missions
GRACE Discharge
Integrating (Geodetic) Data
Soil/Vegetation
Gravity Recovery And
Climate Experiment
since 2003
4
Time variable satellite gravimetry data
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
Gravity field changes induced by mass changes (atmosphere, oceans, hydrosphere)
Observed by GRACE and represented as monthly potential coefficients - Conversion to total water storage anomalies (TWSA)
- Accounting for mass in atmosphere and oceans
m m
300-600 km
filtering
Improve hydrological model with GRACE data
5
WaterGAP Global Hydrology Model (WGHM)
GRACE
0.5°x0.5° grid
- daily time step - monthly averaged potential coefficients
calibration parameters
Döll et al. (2003)
300-600 km
DDK filtered TWSA
w at er co m pa rtm en ts
Uncertainties of GRACE TWSA
6
Standard deviation Error correlations
10 20 30 40
[mm] -1 -0.5 0 0.5 1
[-]
Research gaps
7
Main challenges:
- resolution mismatch
- disaggregation of TWSA
- uncertainties of model and data
- parameters calibration against TWSA
- data-model fusion methods
8
Classical Approach: EnKF
Ensemble Covariance Matrix
Model Prediction GRACE TWSA Observations
) p , u , f(x
x
(1)k
(1)k1…
(1)k1 (1)k1N
ewater state samples at time k forcing
) p , u , f(x
x
(Nk e)
(Nk1e) (Nk1e) (Nk1e)parameters
ε , Ax l
k
kdesign matrix errors
Ne1 i
T k (i)
k k (i)
k e
k
e
(x x )(x x )
1 N ) 1 (x C
Σ
llEnsemble Kalman Filter (EnKF)
1 ll T
k e T
k
e
(x )A (AC (x )A Σ )
C
K
ensemble generation
with gain matrix ) (x C ,
x
k e k
(i)k(i) 1
k
x
x
Evensen (1994) Burgers et al. (1998)
(1)
δl
k Ax
(1)k l
k
(1)
x
kK [ ]
(1)
x
k( )
) (N k
δl
ek
l
) (N k
x
eK [ ]
) (N k
x
e( ) Ax
(Nk e)…
White noise vs.
Correlated errors
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
9
Variant 1: SQRA
SQRA
with gain matrix K as in EnKF
EnKF
Avoid additional source of sampling errors
Update
ensemble mean
Re-initialization required for model perturbations
, C
e(x
k)
k
k
kk
x K l A x
x
(1)
δl
k Ax
(1)k l
kK [ ]
(1)
x
k( )
) (N k
δl
ek
l
) (N k
x
eK [ ]
) (N k
x
e( ) Ax
(Nk e)…
(1)kx
SQuare Root Analysis Scheme (SQRA)
Evensen (2004)
10
Variant 2: SEIK
Pham et al. (1998)
SEIK
1 ll T T
k 1 k 1 ll T T
k 1
k
(G L A Σ AL ) L A Σ
L
K
with
SQRA
1 ll T
k e T
k
e
(x )A (AC (x )A Σ )
C
K
Update in observation space
Singular Evolutive Interpolated Kalman filter (SEIK) Computationally more efficient:
especially, if number of observations >> ensemble size N
eUpdate in ensemble space
#obs x #obs
k
k
kk
x K l A x
x
N
ex N
eAfter some transformations and
k
k
kk
x K l A x
x
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
T k k
k
) L GL
C(x
EnKF results with WGHM
11
Soil water Snow water
Synthetic experiment within the Mississippi River Basin (MRB)
USA
MRB
WGHM GRACE
TWSA EnKF
‘Truth ‘
Schumacher et al. (2016b)
Error correlations of GRACE TWSA
12
4 sub-basins 11 sub-basins 16 grid cells: 5° x 5°
10 – 15 mm 15 – 25 mm 20 – 25 mm
σ σ σ
up to 50 % up to 90 % > 90 %
ρ ρ ρ
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
Root mean square error (RMSE)
[mm]
TWSA
C/DA results: exchange error covariance matrix
13
Correlation of residual
curves* [-]
EnKF USA
MRB
white noise
or correlated errors
Synthetic experiment within the Mississippi River Basin (MRB)
*
after subtracting linear trend, annual and semi-annual cycles
Schumacher et al. (2016b)
WGHM EnKF white
EnKF
correlated
C/DA results: exchange filter algorithm
14
Synthetic experiment within the Mississippi River Basin (MRB)
EnKF or SEIK USA
MRB
correlated errors Root mean
square error (RMSE)
[mm]
Correlation of residual
curves* [-]
*
after subtracting linear trend, annual and semi-annual cycles
TWSA
Schumacher et al. (2016b)
EnKF SEIK
EnKF SEIK
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
Murray – Darling River Basin (MDB)
15
USA
MRB Transfer Australia
Groundwater loss
WGHM
GRACE EnKF
Schumacher et al. (2016c)
w ater com par tme nts
16
Conclusions and Outlook
Insights about data assimilation frameworks:
- EnKF is commonly used due to simple implementation
- Alternative methods show benefit for state / parameter estimation uncertainties, computational efficiency
- effect of implementing GRACE error correlations visible Recommendations:
- consider GRACE error correlations - alternative methods
Further investigations will contribute in enlarging our
understanding of hydrological data assimilation and parameter calibration
- parameter calibration
- constraints for individual water states - disaggregation of GRACE TWSA
- C/DA methodology
1 Motivation 2 Frameworks 3 Applications 4 Conclusions
17
Contact:
Maike Schumacher Astronomical, Physical and Mathematical Geodesy University of Bonn, Germany email: schumacher@geod.uni-bonn.de homepage: http://www.igg.uni-bonn.de/apmg/index.php?id=schumacher
Thank you for your attention!
Any questions?
18
Main References
Ensemble Kalman Filter Approaches
- Evensen (1994) Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics. J Geophys Res 99(C5):10143-10162. doi:10.1029/94JC00572
- Burger et al. (1998) Analysis scheme in the ensemble Kalman filter. Mon Weather Rev 126:1719-1724. doi:
10.1175/1520-0493(1998)126$<$1719:ASITEK$>$2.0.CO;2
- Evensen (2004) Sampling strategies and square root analysis schemes for the EnKF. Ocean Dynam 54:539-560.
doi:10.1007/s10236-004-0099-2
- Pham et al. (1998) A singular evolutive extended Kalman filter for data assimilation in oceanography. J Marine Syst 16(3-4):323-340. doi:10.1016/S0924-7963(97)00109-7
Calibration and Data Assimilation Approach:
- Schumacher M. (2012): Assimilation of GRACE data into a global hydrological model using an ensemble Kalman filter. Master thesis, University of Bonn.
- Eicker et al. (2014): Calibration/Data Assimilation Approach for Integrating GRACE Data into the WaterGAP Global Hydrology Model (WGHM) Using an Ensemble Kalman Filter: First Results. Surv Geophys, Vol. 35 (6), pp. 1285-1309.
doi:10.1007/s10712-014-9309-8
- Schumacher et al. (2016a): A Systematic Impact Assessment of GRACE Error Correlation on Data Assimilation in Hydrological Models. J Geod, doi:10.1007/s00190-016-0892-y
- Schumacher et al. (2016b) Exploring the Hydrologial Millennium Drought Pattern (2003-2009) in the Murray-Darling Basin by Assimilating GRACE into WGHM (in preparation)
WaterGAP Global Hydrology Model (WGHM):
- Döll et al. (2003): A global hydrological model for deriving water availibility indicators: model tuning and validation. J Hydrol, 207, pp. 105-134. doi:10.1016/S0022-1694(02)00283-4
- Müller Schmied et al. (2014): Sensitivity of simulated global-scale freshwater uxes and storages to input data,
hydrological model structure, human water use and calibration. HESS 18:3511-3538. doi:10.5194/hess-18-3511-2014