1 Approximate decorrelation and non-isotropic smoothing of time-variable GRACE gravity field models Jürgen Kusche, Roland Schmidt with input from Susanna.

1

Approximate decorrelation and non-isotropic smoothing of time-variable

GRACE gravity field models

Jürgen Kusche, Roland Schmidt

with input fromSusanna Werth, Roelof Rietbroek

GFZ Potsdam

IUGG 2007, Perugia, GS002

2

Outline of the talk

• GRACE fields exhibit artefacts (“stripes”) which may be seen as a realization of spatially correlated noise - smoothing and/or “de-striping” is required

• Theory: Discussion of ways to decorrelate (“de-stripe”) the noise in GRACE solutions (including method from Swenson-Wahr 2006 (SW06) and a new method)

• Theory: The scaling (bias) problem

• Results: De-striped GFZ GRACE RL4 fields, surface mass grids, and a time series of basin-averaged GRACE-derived OBP ( talk in JGS001)

• Conclusions

3

“Stripes” in GRACE solutions

4

Stripes in GRACE solutions

NS-oriented artefacts

gravity field determination =ill-posed problem

Stochastic (noise) and deterministic (background model) errors cause unphysical oscillations

RMS variability of 40 GFZ RL04 monthly solutions in

2/03-12/06 relative to their mean (7-10/04 and

12/06 excluded), Gaussian 550km

Surface Mass

5

Decorrelation, “de-striping”

6

• degree-dependent (isotropic)• Gauss (Jekeli 1981, Wahr & al 1998), Gauss-Weierstrass (Freeden 1998), Hanning (Jekeli 1981), Blackman (Schmidt & al 2006), CuP (Fengler & al 2006)

• degree- and order-dependent• modified Gauss (Han 2005)• removing single coefficients based on hypothesis testing (Sasgen & al 2005)

• full non-isotropic (general two-point kernel)• constrained fields (Tikhonov)• empirical signal decorrelation combined with Gaussian (Swenson & Wahr 2006)• empirical error decorrelation and Tikhonov smoothing (Kusche 2007)

Issues • de-striping property• amplitude damping (bias) and phase lags• interpretability• optimality criteria, multiresolution properties

Filter Methods for GRACE-L2 Products

7

• combine approximate error decorrelation and Tikhonov smoothing (Kusche 2007)• scaled dense synthetic, “smooth” normal matrix for 1 month• synthetic, smooth signal variance model from Hydrology + Ocean circulation• damping “on normal equation level”

This work

8

Construction of E and S GRACE orbits (coverage)

Hydrology Model +Ocean Circulation Model

9

LAT=60oCross-sections

N-S direction (o)

W-E direction (*)

Impulse response

Filter Properties

This work

LAT=0o

Distance from kernel center

10

This work Swenson and Wahr (2006)

black circle = Gaussian 500km

Impulse response

Filter Properties

11

can be approximated as block-diagonal

Decorrelation/Smooth. Filter W for L = 70

12

C-Block (m+1)

odd/even degrees

Asymmetric order/parity weighting

degree

C-Block (m)

13

Scaling (bias) problem

14

• ratio between filtered and exact basin average

• depends on • filter• shape of basin• signal within and outside basin

All smoothed GRACE-based functionals, global maps or basin averages, are systematically biased low

Scaling (bias) problem

• damping of the global rms

15

Relative bias from true and filtered signal, including hydrology

apparent phase lag

Scaling (bias) problem

16

Relative bias from true and filtered signal, hydrology removed

400km: 56%

Scaling (bias) problem

year

17

Comparison of filters based upon variance and standard scaling bias

Comparison Gaussian – This Work

18

Results

19

Gaussian Filter

RMS variability of 40 GFZ RL04 monthly solutions in 2/03-12/06 relative to their mean (7-10/04 and 12/06 excluded)

Further “de-striping” reduced amplitude (biased towards zero)

Left: Gaussian 500km, Right: Gaussian 550km

wrms=3.85cm

Surface MassGeoid

20

Empirical signal decorrelation according to Swenson and Wahr

(2006)

Filter > l=10, Gaussian 400km

RMS variability of 40 GFZ RL04 monthlySolutions in 2/03-12/06 relative to theirMean (7-10/04 and 12/06 excluded)

wrms=3.76cm

Decorrelation – Swenson and Wahr 2006

Surface Mass

21

RMS variability of 40 GFZ RL04 monthly solutions in2/03-12/06 relative to their mean (7-10/04 and 12/06 excluded)

Left and right: approx. decorrelated using 8/03 orbits and LaD+ECCOfor W-matrix (up to deg/ord = 70), a = 10E+14

wrms=3.83cm

Decorrelation – This Work

Surface MassGeoid

22

BOTH are decorrelated/smoothed using the SAME operator, i.e.8/03 orbits and LaD+ECCO for W-matrix (up to deg/ord = 70), a = 10E+14

Approx. (GRACE) decorrelation does not distort hydrology model

wrms=3.85cmwrms=2.30cm

Decorrelation – This Work

Surface Mass – GRACESurface Mass - WGHM

23

DFG-Mass Transport project STREMP See talk by L. Fenoglio et al in JSG001

Regional Averaging

GRACE “raw” time series of mass change over the Mediterraneanby different methods

24

• Stripes in GRACE solutions still visible; although RL04 improvement over earlier releases

• Best strategy: remove during processing (but perfect de-aliasing impossible)

• Second-best strategy: post-processing using error correlation model (here: from an arbitrary GRACE- or GRACE-type orbit + a-priori model information)

• Proposed technique removed stripes much more effectively compared to Gaussian; simultaneously smoothing (“amplitude bias”) is comparable to Gaussian

• Use for mass transport studies (hydrology, ocean); higher resolution at comparable damping

Conclusions

25

Thank you

26

Decorrelation – This Work

Full W-matrixOrder/parity only

RMS variability of 40 GFZ RL04 monthly solutions in2/03-12/06 relative to their mean (7-10/04 and 12/06 excluded)

Approximately decorrelated using 8/03 orbits and LaD+ECCOfor W-matrix (up to deg/ord = 70), a = 10E+14

27

Decorrelation – This Work

W-matrix based on syntheticnormals from orbit 8/03

W-matrix based on covariancematrix for 8/03 GFZ-RL04

RMS variability of 40 GFZ RL04 monthly solutions in

2/03-12/06 relative to their mean (7-10/04 and 12/06 excluded)

Apriori model information for W-matrix (70,70) from LaD+ECCO, a = 10E+14

28

Decorrelation – This Work

29

Filtering Regularization

eq.

interpretation

estimate of x estimate of x

reduces variance

yes yes

introduces bias

yes (“scaling factor”)

yes

alternative interpretatio

n

unbiased estimate of averaged x

no

required processing

L2 data L1 data (but…)

LSx̂Wx̂

LS

1T11T x̂WyA)RAA(x̂

Decorrelation – This Work

30

Filtering Regularization

parameter tuning

S/N geophysical signals /GRACE

(Wiener/”optimal”)

S/N geophysical signals/ GRACE (LSC)

ordata-driven (GCV,VCE)

latitude-dependence

no (but…) automatically

anisotropic decorrelatio

nno (but…) automatically

geometric interpretatio

nyes no

Decorrelation – This Work

31

Spherical disc signal + Gaussian(can be analytically treated)

Disc radius [km]

Gauss

ian

sm

ooth

ing r

adiu

s [k

m]

amplitude scaling error (relative bias)Mediterranean

Amplitude Scaling Error - Gaussian