characteristics of input features are presented. Because the
results from
WBNN
and
BPNN
are not very different, this type of
information is apparently captured sufficiently well using the
more basic
BPNN
. Since the results for each scenario and the
AOI
are very similar between the
BPNN
and the
WBNN
algo-
rithms, detailed plots of the results are only presented for the
BPNN
in the following discussion.
Figure 8 shows the results after applying the algorithm
based on scenario 1 to the four areas of study. The plots show
the relationship between the testing data estimates from the
algorithm and the target data which is the
EC
reference data.
The correlation between the
EC
estimates and targets for the
test data is indicated by the correlation coefficient
R
. As
Figure 8 shows, the correlation between the estimation and
target in
AOI
1 is weaker than for the other areas. This is con-
sistent with Equations 3, 4, and 5 where if the height of grass
increases,
σ
0
veg
and volume scattering increase, and therefore
we get a weaker relationship between the soil moisture and
the backscatter coefficients. A better estimate of conductiv-
ity is obtained in
AOI
3 (see Figure 2c) which covers a mixed
bare and weed area. It seems this bare soil provides stronger
correlation between the
SAR
data and conductivity. This is ex-
pected due the fact that more vegetation will contribute a rela-
tively greater amount to the backscatter than the soil it covers,
thus resulting in less correlation to the actual soil moisture.
By comparing the results of
AOI
2 and
AOI
4, it can be inferred
that the dead grass (laid over short Bermuda grass) reduces
the estimation correlation, perhaps due to the increased atten-
uation and the presence of higher volume scattering. As can
be seen from Table 2,
AOI
4 has weaker correlation than
AOI
2
by 15 percent, which may be due to high volume scattering.
Since each area has a different conductivity range, we did not
compare the
RMSE
between them. We use the
RMSE
evaluation
for comparing the scenario results within each
AOI
.
Figure 9 shows the results of applying the algorithm to
the area of study based on scenario 2. As the figure shows,
inclusion of the mean feature for sliding window (3
×
3)
improves the correlation between the estimates and the target
for all
AOIs
. Since this feature is a texture feature (when used
in combination with the individual pixel backscatter values),
we expect to have improvement in the estimation based on
Equations 6, 7, and 8. This is because some knowledge of the
surface roughness gained from the texture feature can im-
prove the NN’s ability to estimate the dielectric constant (
ε
)
and thus also
EC
. The amount of correlation improvement is
around 10 to 20 percent in all
AOIs
when this texture feature
is included. Therefore, we can conclude that using a mean of
HH
and
VV
around a pixel can improve the soil conductivity
estimation. Figure 9 also shows that
AOI
3 has better correla-
tion than other
AOIs
in this scenario.
As we compare
RMSE
and correlation results of scenario
2 to scenario 1 in Table 2, we can find correlation improve-
ment at all
AOIs
of 8 to 18 percent. Also, the
RMSE
is lower in
scenario 2.
RMSE
decreases more in
AOI
3 in which we have
some bare soil.
Figure 10 shows the result of conductivity estimation
based on scenario 3. By comparing this output to the results
of scenarios 1 and 2, we see more improvement for all
AOIs
.
Similar to scenarios 1 and 2, the best estimate of conductiv-
ity is obtained in
AOI
3 where a correlation of approximately
64 percent is obtained. By comparing the results of scenarios
2 and 3 in Table 2, we can see that scenario 3 improves the
correlation for most
AOIs
with a 1 to 17 percent improvement.
The results also show that scenario 3 has stronger correlation
Figure 8. Conductivity estimation for the AOIs in scenario 1 [using only SAR backscatter coefficients of HH and VV along with the local
incidence angle (
θ
)]
516
July 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
