133
L
L
km. L § low
;velopment of the
10ut the growing
:ypes. Although
Intensity scales,
se measurements,
efficient of the
le images and
neasurements,
late and in the
factor was applied
:e 4 is for
Lng angle. The
) the calibration
lie. The radar is
ilibration. The
> preprocessing
care of the
irrection and
calibration
letter, as was
:ors.
i large contrast
Figure 4. Development of radar backscatter throughout
the growing season for some croptypes. X-band SLAR,
horizontal polarization, 15° grazing angle.
exists between winterwheat and the other croptypes in
April and May. In June the contrast is very small,
while all the crops are in their growing stage. In
July a good contrast is present between all the crop-
types, whereas in August the development of the
backscatter coefficient of potatoes interferes with
the one for winterwheat.
The large contrast between winterwheat and the other
croptypes in the early growing season only exists at
low grazing angles. It can be explained as follows:
the wintercrops, like winterwheat, are planted before
winter and start growing in this area in April. The
other croptypes are planted in April and May and show
their biomass not before the end of May. Although the
ground coverage by the new plants is small, the
backscatter at low grazing angles is increased,
because the smooth soil alone gives a very small
amount of backscatter at these angles, so the small
leafs sticking out of the ground contribute consider
ably to the total backscatter. At larger grazing
angles say around 40°, the backscatter from the fields
is much increased and the previously described effect
is smaller, resulting in very little to no contrast
between these croptypes.
Thus we should be able to distinguish between
winter- and summer crops from one flight in April or
May, and since our testarea contains mainly one
wintercrop, namely winterwheat, we should be able to
identify all winterwheat fields. Figure 5 shows the
histogram of the field averaged radar backscatter
coefficients of the SLAR image from May. From this
figure it is clear that the winterwheat fields can
be completely separated from the other fields, simply
by applying a threshold level.
Now that the winterwheat is identified, we must try
to classify the remaining fields from other flights.
This demonstrates the hierarchy in our classifier in
contrast with the previous classification experiment
(Ref. 1) where the time dependence of the radar back
scatter throughout the growing season was used as
discriminator.
Sofar the design of the classifier was straight
Figure 5. Histogram of the May flight: winterwheat
(right) is separated from the other croptypes.
forward and rather simple. However to derive an
optimum result more elaborate methods should be used
to investigate the data. Our main purpose is to make
a selection from the available features per field.
Eigenvalue or principal component analysis can be used
to reduce the dataset into a set of uncorrelated
features. This is done by a dataprojection on two or
more Eigen vectors, which are determined from the
covariance matrix of the dataset.
An evaluation of the dataset using this method
showed that the first two eigen vectors contained
91% of the total variance, which means that the other
four eigen vectors may be deleted. The first eigen
vector is mainly determined from the April- and May
features, whereas the second eigen vector is in fact
a combination of the two July features, so the two
flights at different altitudes.
Since the datasets from April and May are highly
correlated (correlation coefficient 0.91), the dataset
of May was chosen as before and furthermore we
selected the two July features. Figure 6 shows feature
space plots for May versus July and for the 2 July
features. A cross reference of the labels used in this
and other figures can be found in Table 1. In both
plots clusters of croptypes can be distinguished.
A projection on one of the axes makes the classes
inseparable, except for the wheat in May of course and
the sugarbeets in July. The combination of the two
July features means that we deal here with angular
dependences to obtain discrimination. The short time
interval between these two measurements more or less
guarantees that the differences are only caused by
the change in incidence angle. Therefore the clusters
croptype
label
label
potatoes
sugarbeet
winterwheat
peas
onions
oats
winterbarley
beans
grass seed
spinach
A
B
T
E
U
H
GR
BO
GZ
SP
1
2
3
4
5
6
7
8
9
10
Table 1. Legend to plotlabels
