Using water storage variations derived from GRACEto calibrate a global hydrology model

S. Werth, A. Güntner, S. Petrovic, R. Schmidt, J. Kusche

GeoForschungsZentrum (GFZ) Potsdam, Germany

Workshop Hydrology from Space, Geneva 11/2007

Background

Calibrationof the global hydrological model WGHM

Water balance of a river basin:

P = E + R + ΔS

P: PrecipitationE: EvapotranspirationR: RunoffΔS: Water storage change

Model input

Calibration variable

(measured time series of river discharge)

Objectives

Multi-criterial calibrationof the global hydrological model WGHM

Water balance of a river basin :

P = E + R + ΔS

P: PrecipitationE: EvapotranspirationR: Runoff (measured time series of river discharge)ΔS: Water storage change (basin-average values from GRACE)

Calibration variables

Model input

The WaterGAP Global Hydrology Model (WGHM)

• Conceptual waterbalance model

• 0.5° spatial resolution

• Daily time-step

• Human water use accounted for

• Climate forcing data from CRU, GPCC, ECMWF

• Calibration for river dischargeat 1200 stations worldwide

Total continental storage change:

ΔS = ΔScanopy + ΔSsnow + ΔSsoil + ΔSgw + ΔSlakes + ΔSwetl + ΔSriver

1) Analyse model properties with respect to storage change

- Identify sensitive model parameters- Uncertainty assessment

2) Select adequate GRACE data and filter tools

3) Perform multi-objective model calibration

5

Work steps

Hydrological model calibration with GRACE data

Step 1.1) Selection of calibration parameters

Hydrological model calibration with GRACE data

Parameter sensitivity varies regionally with the dominant water storage processes / components

Step 1.2) Model uncertainty analysis

Hydrological model calibration with GRACE data

Structural model errors may exist if model uncertainty range does not enclose GRACE data

1) Analyse model properties with respect to storage change

- Identify sensitive model parameters- Uncertainty assessment

2) Select adequate GRACE data and filter tools

3) Perform multi-objective model calibration

5

Work steps

Hydrological model calibration with GRACE data

Step 2) Selection of adequate GRACE filter tools to derivebasin-average water mass variations

Hydrological model calibration with GRACE data

Filter type Parameter for filter intensity Source

Gaussian filter (GF) filter width Jekeli, 1981

Filter optimized for basin shape (OF) max. satellite error Swenson and Wahr, 2002

Filter optimized for exp. signal model (MF) correlation length, signal variance Swenson and Wahr, 2002

GRACE signal-noise-ratio optimized (SF) factor of formal errors Seo et al, 2005

Correlation Error Filter (CEF) filter window properties Swenson and Wahr, 2006

Decorrelation Filter (DDK) covariance matrix parameter Kusche, 2007

Step 2) Selection of adequate GRACE filter tools: Amazon basin

Hydrological model calibration with GRACE data

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0Pe

rfom

ance

valu

e

500 1000 1500filterwidth (km)

5101520max. satellite error (mm)

DDK decorrelated

w =10, n =30a e

w=30, n =30

a

e

w=20, n =30

a

ew =10, n =30

a

e

w =20, n =30a e

w =30, n =30a e

GFCEF decorrelated

OFCEF decorrelated

0

w =30, n =30a

ew =20, n =30

a

e

w =10, n =30a

e

SFCEF decorrelated

5 10 15 20error factor

50

100

150

2004006008001000corrlation length (km)

= 250 mmσs

w =30, n =30a e

w =20, n =30a e

w =10, n =30a e 20

MFCEF decorrelated

Gaussian filter (GF)

Filter optimized for basin shape (OF)

Filter optimized for exp. signal model (MF)

GRACE signal-noise-ratio optimized (SF)

Correlation Error Filter (CEF)

Decorrelation Filter (DDK)

Step 2) Selection of adequate GRACE filter tools: Lena basin

Hydrological model calibration with GRACE data

Gaussian filter (GF)

Filter optimized for basin shape(OF)

Filter optimized for exp. signal model (MF)

GRACE signal-noise-ratio optimized (SF)

Correlation Error Filter (CEF)

Decorrelation Filter (DDK)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

500 1000 1500 5101520

w=10, n

=30

a

e

w=30, n

=30

a

e

w=20, n =30

a

e

w =30, n =30a e

w =20 , n =30

a

ew =10, n =30a

e

CEF decorrelated CEF decorrelated

DDK decorrelated

CEF decorrelated

0 5 10 15 20

w =20, n =30a e

w =10, n =30a e

w =30, n =30a e

CEF decorrelated

2004006008001000

w =30, n =30a e

50

250= 300 mm

200

150

100

20

Perf

oman

ceva

lue

filterwidth (km) max. satellite error (mm)

GF OF SF

error factorcorrlation length (km)

MF

Step 2) Selection of adequate GRACE filter tools

Hydrological model calibration with GRACE data

1014, 4

30, 3, 2, 30

250, 200

0.4

13

200

Parameter Value

a, p

wa, we, na, ne

σs [mm], cl [km]

error factor

εmax [mm]

rg [km]

Parameter

0.74

0.63

0.78

0.77

0.78

0.75

Performance valueExample: Amazon basin

FilterGaussian filter (GF)

Filter optimized for basin shape (OF)

Filter optimized for exp. signal model (MF)

GRACE signal-noise-ratio optimized (SF)

Correlation Error Filter (CEF)

Decorrelation Filter (DDK)

DDKGFMississippiSFSFVolgaMFCEFMFYukon

MFMFGangesOF,MFOF, MFAmazonGLDASWGHMBasin

Optimal filter for 5 example river basins and two global hydrological models

1) Analyse model properties with respect to storage change

- Identify sensitive model parameters- Uncertainty assessment

2) Select adequate GRACE data and filter tools

3) Perform multi-objective model calibration

5

Work steps

Hydrological model calibration with GRACE data

Ob basin

1.0 0.5 0.0 -0.51.0

0.5

0.0

-0.5

-1.0

-1.5

-2.0

Nash-Sutcliffe coefficient for water storage change

Nas

h-S

utcl

iffe

coef

ficie

nt

for r

iver

dis

char

ge

6

perfect model

simulation

Hydrological model calibration with GRACE data

Step 3.1) Test for single-objective calibration

WGHM Monte-Carlo run

Standard WGHM calibrated for river discharge

WGHM single-objective calibration for GRACE storage change

Evaluationof error

stop ?

parameter set ranking

Optimalsolution

yes

no

initial parameter

sets

Model simulation

GRACE total storage

variation

RunoffMeasurement

data

7

0

erro

rdi

scha

rge

Pareto Frontier

0

single model simulation

error total storage change0

Hydrological model calibration with GRACE data

Step 3.2) Multi-objective calibration approach

Parameter variations

Model simulation

DDS Dynamically Dimension Search► single-objective calibration algorithm extended for mutli-objective problems

NSGA-II Non-dominated Sorting Genetic Algorithm► evolutionary multi-objective calibration algorithm

14

Hydrological model calibration with GRACE data

Step 3.3) Implementation of multi-objective calibration algorithms into WGHM

14

Hydrological model calibration with GRACE data

Conclusions

• Model parameter sensitivity for water storage change varies regionally with the dominant storage components

• GRACE data help to identify structural model errors during model uncertainty assessment

• Adequate filter tools to derive basin-average storage change from GRACE vary with location and river basin characteristics

• GRACE data are highly valuable to constrain large-scale hydrological models in a multi-objective calibration framework

Using water storage variations derived from GRACEto calibrate a global hydrology model

S. Werth, A. Güntner, S. Petrovic, R. Schmidt, J. Kusche

GeoForschungsZentrum (GFZ) Potsdam, Germany

Workshop Hydrology from Space, Geneva 11/2007