370 likes | 481 Views
Realistic error modelling for InSAR Tim J Wright, Gareth J Funning (COMET, Oxford Earth Sciences) Peter J Clarke (Geomatics, University of Newcastle) Charles W Wicks (USGS, Menlo Park, California). Components of interferometric phase.
E N D
Realistic error modelling for InSARTim J Wright, Gareth J Funning (COMET, Oxford Earth Sciences)Peter J Clarke (Geomatics, University of Newcastle)Charles W Wicks (USGS, Menlo Park, California)
Components of interferometric phase Dfint = Dfgeom +Dftopo +Dfatm + Dfnoise + Dfdef Dfint = Dfgeom +Dftopo +Dfatm+ Dfnoise + Dfdef A foggy morning, near ancient Mycenae, Greece
June to December July to December June to July Components of interferometric phase Dfint = Dfgeom +Dftopo +Dfatm + Dfnoise + Dfdef Dfint = Dfgeom +Dftopo +Dfatm+ Dfnoise + Dfdef Turbulent atmosphere Athens Earthquake – September 1999
Comparison with Rain Radar (The Netherlands) Figures courtesy Ramon Hanssen (Delft)
The “Dutch Alps” Figures courtesy Ramon Hanssen (Delft)
17 August 1999 Izmit (Turkey) earthquake ERS2 – ERS2 13-Aug-1999 – 17 Sept-1999 ERS1 – ERS1 12-Aug-1999 – 16 Sept-1999
mm 84 0 -84 • ERS1 interferogram – ERS2 interferogram • Atmosphere + Orbital Errors.
Solutions to atmospheric water vapour problem? • Ignore (most common) • Quantify • Model based on other observations • (e.g. GPS, meteorology…?) • Increase SNR by stacking • For some it is the signal not the noise.
Solutions to atmospheric water vapour problem? • Ignore (most common) • Quantify • Model based on other observations • (e.g. GPS, meteorology…?) • Increase SNR by stacking • For some it is the signal not the noise.
Question. • How can we quantify the influence of spatially-correlated noise on our models?
Monte Carlo Simulation of Correlated Noise A. Determine an empirical Variance-Covariance Matrix (VCM) for the noise in the interferogram. B. Create a suite of pseudo-random realisations of the noise, and add to the original data. C. Invert each dataset – the range of values for each model parameter gives their error.
A: Create a VCM – a practical approach • Calculate the covariance function - chose part of image away from signal or remove a 1st pass model. - remove plane, set mean to zero. - covariance function is radially-averaged autocovariance. - this can be calculated spatially or in the frequency domain (e.g. Hanssen, 2001).
radians radians 10 km 10 km Covariance / radians2 20 km Distance / km A: Create a VCM – a practical approach
sub-image size variance length scale 2 10 km 0.24 rad 1.6 km 30 2 25 km 1.1 rad 4.8 km 2 50 km 5.2 rad 10 km 2 100 km 28 rad 24 km 2 160 km* 30 rad 33 km 20 *= whole image 10 0 -10 0 20 40 60 80 100 120 Distance / km A word of caution…
A: Create a VCM – a practical approach • Calculate the covariance function • Try functional fit: e.g. cov = s2 e-ar cov = s2 e-ar cos(br) [s2= variance, r = distance] 3. Use this function to construct VCM for the sampled points of the interferogram.
B. Create synthetic noise • Construct X[vector of Gaussian noise, mean 0, s 1] • Want Y[vector of correlated noise with known VCM Sy] • Let Y = LXSy = LSxLT[L is a matrix] • But Sx= I Sy = LLT • i.e. L is the Cholesky Decomposition of Sy
C. Invert perturbed datasets • Chose your favourite inversion method!
Test – correlated vs independent noise • Grid denser than 5km: independent noise error estimate too small • correlated noise “correct” error estimate • Grid sparser than 5km: both methods overestimate error.
Slip (m) 4 3.5 80 100 Dip Example 1: Fault geometry
25 Volume (10-3 km3) 15 3.5 5.5 Depth (km) Example 2: Mogi Models Three Sisters: ‘clean’ (8/96 to 10/00) Three Sisters: ‘noisy’ (8/97 to 9/00)
20 km Example 3: Distributed Slip (Izmit Earthquake) Dip: 87s 88s 86s 88n 81n 61n Rake: 174 171 178 -178 -164 -168 6-segment model (Wright et al, 2001) r.m.s. misfit = 28 mm; M0 = 265 x 1018 Nm
20 km Example 3: Distributed Slip (Izmit Earthquake) 6-segment model (Wright et al., 2001) r.m.s. misfit = 28 mm; M0 = 265 x 1018 Nm Geometry of Wright et al., but with variable slip (5km patches) r.m.s. misfit = 24 mm; M0 = 260 x 1018 Nm Vertical fault, pure strike slip (5km patches) r.m.s. misfit = 45 mm; M0 = 288 x 1018 Nm
20 km Example 3: Distributed Slip (Izmit Earthquake) 25 km 6-segment model (Wright et al., 2001) r.m.s. misfit = 28 mm; M0 = 265 x 1018 Nm Geometry of Wright et al., but with variable slip (5km patches) r.m.s. misfit = 24 mm; M0 = 260 x 1018 Nm Difference between 6-segment model and variable slip model Vertical fault, pure strike slip (5km patches) r.m.s. misfit = 45 mm; M0 = 288 x 1018 Nm
Example 3: Distributed Slip (Izmit Earthquake) Slip: Error:
Model Covariance 2 km 6 km 10 km 14 km 18 km 22 km
Inversion Method Elastic structure Fault Geometry EARTHQUAKE SLIP MODEL Smoothness criteria Data sampling scheme Data Noise
Mitigation using additional data • InSAR alone Simulations using one possible configuration of a dedicated InSAR satellite
Mitigation using additional data • InSAR alone Black dots – 1 int (descending, right-looking) Red dots – 2 ints (asc + dsc, right-looking) Cyan dots – 3 ints (asc + dsc, right + left-looking)
Mitigation using additional data 2. GPS data slip error
InSAR Data Only Long Period Seismology Data Only InSAR and seismology Mitigation using additional data 3. Seismic Data
Conclusions • Spatially-correlated noise must be considered when determining errors. • The method presented is relatively simple to use, and can be applied to any model derived from InSAR data. • Earthquake slip distributions derived from geodetic data are highly non-unique. • Use of additional data can often dramatically reduce parameter errors and trade-offs
Components of interferometric phase Dfint = Dfgeom +Dftopo +Dfatm + Dfnoise + Dfdef Dfint = Dfgeom +Dftopo +Dfatm+ Dfnoise + Dfdef Layered atmosphere 29/8/1995 to 29/7/1997 30/8/1995 to 29/7/1997 Topography