rovided
ixel po-
jendent
the dis-
loped a
at each
ne flight
ated out
rmation
with re-
ed, the
eorefer-
1 speckle
6). This
the data
isolated
ne back-
ameters
n object-
ighness,
he influ-
! regions
surface
hus, the
ly of the
d param-
ording to
arders of
ughness
s. Based
asses for
roperties
speckle.
Iso avoid
over 200
east 500
side the
a series
. The re-
shown in
; inside a
f the soil
moisture. This is subject of ongoing research and will be
presented in a following paper.
3 REGRESSION ANALYSIS METHODS
3.1 Signal-to-soil moisture correlations
In order to investigate the interdependencies between the
soil moisture measurements and the pixel values of the
SAR image, we measured the correlation between them.
As a first try, we have made a regression analysis for all the
measurements from both campaigns. The correlation coef-
ficients observed this way are listed in the first and fourth
column of table 1.
(0-4 cm) TDR
17th.+21th. 17th. 21th. [17th.+21th. 17th. 21th.
Ohh ] 038 — 27 941 1 08 08 06
cw Í o4 — 027 fae 031 ,043 043
Lhh | 05 -007 00 | 05 047 05
tw | 048 03 06 | 041 041 0
Uw 1.05 02 072] 059 . 04 og
Lvh | 05 038 066 | 05 047 065
L2hh -010 -010 0,46 0,46
L2w 0,08 0,08 0,39 0,39
L2hv 0,22 0,22 0,38 0,38
L2vh 0,16 0,16 0,37 0,37
Phh 0,62 0.31.4 0.76 0.43 049 . 047
Phy -089 001—(0:39 0,12 03 0,37
Table 1: Correlations between radar return and soil mois-
ture.
In general, the correlations obtained for the data from April
21th are much better then the ones for the 17th. On 17th
of April, the E-SAR acquired two measurements with the L-
band (named L and L2). Table 1 shows that the two L-band
measurements behave differently. One observes, that the
local incidence angle plays an important role and can even
predominate over the influence of soil moisture on the pixel
values.
The best correlation was found between the C-band pixel
values and the soil moisture measurements in the 0-4cm
layer, namely: (r = 0.63). Also for the L- and P-bands,
this layer leaded to the best correlations (r = 0.58 for L-
VH and r — 0.62 for P-HH). We note, that the L2-band
values correlate a little better with the TDR measurements
than with the measurements in the 0-4cm layer: (r = 0.40
against r zz 0.09). This is what we have expected, since the
penetration depth depends on the incidence angle.
The columns 2, 3 and 5, 6 in table 1 show the correlation
values for each campaign. The incidence angles at the first
campaign are higher than those at the second (except for
the first L-band, which is even smaller). Anyway, the differ-
ence of 3 degrees is in our opinion to small to explain the
correlation differences.
To analyse the results for each lot (Fig. 3) , we compared
the TDR measurements with the backscatter coefficients in
the C-VV image (in dB). We have used the TDR measure-
ments because the number of soil samples taken in the
0-4cm layer is to small to pretend of a reliable regression
analysis. However, considering that a linear relation be-
tween the TDR measurements and the gravimetrical mea-
surements in the 0-4cm layer (r = 0.92) exists, the correla-
tions with the TDR-values should describe the phenomenon
as well.
17.04.1997 : correlations related to the fields
#1: Schafer
8i 2: Schiling
A 3. Fränkie
e 4: Häberle
x 5: Keller
0 6: Specht
+ 7: Intensiv.
Backscatter coefficient C-VV (dB)
me TOR (Vol-%)
Figure 3: Correlations between C-VV backscattering coef-
ficient and the TDR-measurements.
The best correlation was found for the lot Haberle (with
r = 0.96), which is a homogeneous almost bare field.
The overall tendency is in agreement with the expected be-
haviour. It shows an increasing backscatter coefficient with
increasing wetness, although this is less reliable for the lots
covered with vegetation.
The regression analysis was performed both with the EPOS-
filtered data and with the dataset smoothed within object
boundaries. No worth to mention differences did occur.
However, the method of smoothing within object bound-
aries has the advantage to consider a region around each
pixel, and consequently it works with more representative
samples.
3.2 Principal component analysis
Principal component analysis of the SAR data in different
bands allows to determine which bands and frequencies
contain the most information, and consequently to reduce
the amount of data.
If the soil moisture has any influence on the radar returns, it
should appear in one principal factor. This way, we looked
for a possible existence of a linear combination between
the different frequencies and the measured soil moisture or
other hydrological parameters.
The first step consisted to scale the data, because the back-
scatter coefficient increases with frequency, and thus the
greyvalues in the P-band images are imbalanced compared
to the C-band values. The best correlation coefficients oc-
curred for the fifth principal factor and the measured soil
moisture (r — 0.88 and r — 0.70) in the first and second soil
layer (4-8cm resp. 0—4cm). The most influencing factors
are P-HH, P-HV and L-VH. The P-band plays an important
role and thus will probably be promising in the soil mois-
ture determination. Unfortunately, this factor occupies the
fifth place and describes only 396 of the whole variance. It
signifies, that the soil moisture does probably not represent
the most important information in the radar images.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 551
