soil moisture retrieval.pptx
TRANSCRIPT
Soil moisture retrieval
over bare surfaces
using time-series radar
observations and a
lookup table
representation of
forward scattering
Seung-bum Kim, Shaowu Huang*, Leung Tsang*,
Joel Johnson**, Eni Njoku
Jet Propulsion Lab., California Inst. Technology
* Univ. Washington
** The Ohio State Univ.IGARSS 2011, Vancouver, Canada
S.B. Kim - 1 -IGARSS 2011
Objectives
• A ‘bare and sparsely vegetation surface covers 13% of the world’s land surface.
• Over vegetated surfaces, a soil moisture signal comes from the bare surface.
– Accurate soil moisture retrieval over the bare surface forms the basis of global soil moisture
retrieval
• We focus on the soil surface where the roughness has the isotropic random distribution.
• We study the retrieval of soil moisture at L-band at 40deg incidence angle to apply to the Soil Moisture
Active Passive (SMAP) mission
• NASA’s SMAP mission
– Global, high-resolution mapping of soil moisture
(top 5cm) and its freeze/thaw state
– Three year mission, due for launch in 2015
– 1000 km-wide swath, enabling 2−3 day revisit
– Retrievals with radiometer (36km resolution), SAR
(3km), and SAR/radiometer combined (9km)
– Multi-pol (HH, VV, HV)
– To achieve the global coverage, the number of
single looks was compromised (60, worst case),
leading to somewhat large speckle. Total radar
measurement error ranges from 0.5dB (13%) to
0.7dB (17%)
S.B. Kim - 2 -IGARSS 2011
Issues with bare surface soil moisture retrieval
• Four parameters dominating the radar scattering from bare surface: surface roughness (s), soil moisture
(Mv), correlation length of roughness (l), correlation function (F)
• Issues to resolve for accurate soil moisture retrieval
– With HH, VV, HV, we cannot determine all 4 unknowns
– Knowledge of correlation length and function is very inaccurate (50% error).
– Ambiguity: dry & rough soil σ0 ~ wet & smooth soil simga0 can lead to multiple solutions
– 13% to 17% radar measurement error
• Past literature
– Empirical retrieval model without considering l or F (Oh et
al. 1992; Dubois et al. 1995)
– The exponential function fits data well (Shi et al. 1997;
Mattia et al. 1997)
– Estimate s and l after assuming temporally static
• Using ancillary dry condition (Rahman et al. 2008)
• Using one time measurement (Joseph 2008)
• Statistical estimate (Verhoest et al. 2007)
• Using ancillary weather model (Mattia et al. 2009)
– Eliminate s during retrieval (Shi et al. 1997)
• Our goal: develop non-empirical and simple method that does
not need ancillary information
S.B. Kim - 3 -IGARSS 2011
Soil moisture retrieval method
• A forward model (Numerical Maxwell Model in 3-Dimension, NMM3D, Huang et al. 2010; 2011)
is inverted using its lookup table representation.
– The NMM3D computes numerical solutions of Maxwell’s equations without approximate
parameterizations or tuning.
– The NMM3D predictions compare well with in situ datasets representing wide ranges of
roughness, soil moisture, and correlation length.
Results with correlation length/rms height (l/s) =4,7,10 is also available.
rmse=1.49dB (VV)
1.64dB (HH)
Huang, Tsang,
Njoku, Chen,
TGRS, 2010
S.B. Kim - 4 -IGARSS 2011
Parameterized IEM based
on 3 Mv and 3 ks simulations
RMS Error = 0.0067 cm3/cm3
Lookup table
RMS Error = 0.0016 cm3/cm3
Courtesy: J.J. van Zyl
10log
shh
= -20.17 +15.33mv+13.63log ks( )
10log
svv
= -18.81+ 25.33mv+10.99log ks( )
• The retrieval based on a lookup table performs better. In comparison, the inversion of a sophisticated model (IEM) is not as accurate.
Retrieval with a lookup table
S.B. Kim - 5 -IGARSS 2011
Soil moisture retrieval method
• Two-polarization (HH and VV) are used for input.
HV channel is set aside for vegetation information
for future. Mv and roughness are retrieved.
• 2N independent input (HH and VV, N is the
number of time-series)
– N+1 unknowns assuming roughness does
not change in time.
• The retrieval of the soil moisture is accomplished
by the least square minimization.
• Time-invariant roughness is estimated first.
• Then time-varying soil moisture is retrieved.
C s,er1,er2 ,..,erN( ) = w1,HH sHH
0 t1( ) - s HH ,LUT
0 s,er1( )( )2
+ w1,VV sVV
0 t1( ) - sVV ,LUT
0 s,er1( )( )2
+w2,HH s HH
0 t2( ) - sHH ,LUT
0 s,er2( )( )2
+ w2,VV sVV
0 t2( ) - sVV ,LUT
0 s,er2( )( )2
+...
+wN ,HH s HH
0 tN( ) - s HH ,LUT
0 s,erN( )( )2
+ w1,VV sVV
0 tN( ) - sVV ,LUT
0 s,erN( )( )2
=1
N[E1 s HH
0 t1( ),sVV
0 t1( ), s,er1( ) + E2 s HH
0 t2( ),sVV
0 t2( ), s,er2( ) + ...
+EN sHH
0 tN( ),sVV
0 tN( ), s,erN( )]
S.B. Kim - 6 -IGARSS 2011
Monte-Carlo simulation
• The radar measurement noise
(0.7dB, 17%) is modeled by a
Gaussian random process.
• Errors in roughness estimate
is smaller than 10%.
• Real part of the dielectric
constant (εr) is retrieved first
and Mv error is smaller than
0.06 cm3/cm3.
Time-series search of lookup table
Snap-shot search of lookup table
• A snapshot search of the
same lookup table has no
constraint on roughness, is
subject to the ambiguity, and
the rmse is very large.
S.B. Kim - 7 -IGARSS 2011
Validation with in situ data
• Truck-mounted radar measurements in
Ypsilanti, Michigan were obtained over
a two-month campaign (Oh et al. 2002).
• The LUT time-series and snap-shot
performs comparably most likely
because the radar measurement error
(~0.4dB) is smaller than 0.7dB.
• Dubois method has outliers.
S.B. Kim - 8 -IGARSS 2011
Validation with in situ data
• The retrievals from all 4 sites are
combined in one scatter plot.
• Many retrievals of the Dubois retrieval
become outliers.
• Overall the LUT time-series retrieval
shows the best correlation with the in
situ mv and the best rmse in soil
moisture estimation.
LUT snapshot
0.0 0.1 0.2 0.3 0.4in situ (cm3/cm3)
0.0
0.1
0.2
0.3
0.4
retr
iev
al (
cm3/c
m3) rmse=0.055
mean_e=0.006
corr=0.82
LUT time-series
0.0 0.1 0.2 0.3 0.4in situ (cm3/cm3)
0.0
0.1
0.2
0.3
0.4
retr
iev
al (
cm3/c
m3) rmse=0.044
mean_e=0.013
corr=0.89
Dubois
0.0 0.1 0.2 0.3 0.4in situ (cm3/cm3)
0.0
0.1
0.2
0.3
0.4re
trie
val
(cm
3/c
m3) rmse > 1
mean_e > 1
corr= -0.14
Relative change index
0.0 0.1 0.2 0.3 0.4in situ (cm3/cm3)
0.0
0.2
0.4
0.6
0.8
1.0
ind
ex
corr=0.80
S.B. Kim - 9 -IGARSS 2011
Effects of correlation length
• The validation with the Michigan data is performed: (a) uses l (correlation length) / s (rms height) of
10 (= the same as forward model). (b) uses the truth. The choice of l/s does not affect the Mv
retrieval.
Site (rms hgt) l/s =4 l/s = 7 l/s = 10 l/s = 15
1 (0.55cm) 0.037 0.041 0.038(a) 0.038(b)
2 (0.94cm) 0.043 0.048 (b) 0.051(a) 0.048
3 (1.78cm) 0.043(b) 0.043 0.047(a) 0.055
4 (3.47cm) 0.055(b) 0.042 0.040(a) 0.042
• Even if the value of the cost function changes, the
location of the minimum (=Mv retrieval) does not.
Retrieved Mv
• Change of l/s merely adds a bias to the cost function
+..
S.B. Kim - 10 -IGARSS 2011
Effects of correlation length
l/s=15
l/s=4
• If a single scattering process is
dominant, the IEM shows that the
roughness effect and the dielectric
effect can be decoupled (Fung et
al. 1992; Shi et al. 1997):
σ0 (Mv, s, l/s) = f(Mv) + g(s, l/s)
Then
σ0 (Mv, s, l/s=a) - σ0 (Mv, s, l/s= b )
= g(s, l/s=a) - g(s, l/s=b)
= independent of Mv.
• The bias offset is uniform wrt
dielectric constant and
polarization.
S.B. Kim - 11 -IGARSS 2011
Summary
• Findings
– Time-invariant roughness and soil moisture are estimated using the time-series method (2N inputs
solve N+1 unknowns).
– Does not require ancillary information.
– Simple search of a lookup table.
– Tested with in situ data: the error is 0.044 cm3/cm3 using 6-11 time-series inputs.
– The time-series method performs better than the other methods.
– Retrieval is mostly independent of the knowledge of correlation length one degree of freedom is
reduced.
• Discussion
– Increase in time-series window will further reduce the radar measurement noise improve
roughness estimate improve Mv retrieval.
S.B. Kim - 12 -IGARSS 2011
backup
S.B. Kim - 13 -IGARSS 2011
Soil moisture: retrieval performance