383
as were: 2nd
th April, 18th
. All are path
jne 1980 which
the use of
d register the
her image to
sary for pixel
d another. The
multitemporal
le accuracy of
s, (Anderson,
nto ensuring a
nsforming the
out using the
age on a GEMS
IME 2250. This
ts with known
cale O.S. maps
coordinates to
rm equations,
icinity of the
6 scenes, all
s with average
100m for each
ch image was
erpolation was
interpolation
cally correct
ting, 603000,
sting, 653000,
from each date
nd haze in the
re degradation
data suffered
ult to remove.
false colour
el by 512 line
maximum ground
to the imagery
ng coordinates
ata into X and
on proceedures
used to locate
. The training
or all images,
ebruary, April
on the August
ene meant that
stablished for
new areas were
re removed. In
, one for each
wheat, winter
sugar beet,
us woodland, a
dsat data are
in this study,
nds for crop
process was
decision tree
ms for each of
by using the
'hese data were
and plotted as
-indenlaub and
The coincident spectral plots show comparative
changes in the temporal spectral response of all the
cover types and could be used to identify a subset of
bands / dates for optimum cover type separation
through the use of decision boundaries. This involves
the identification of numerical limits within the
coincident spectral plots for the separation of cover
types. In this case where the data has both a
spectral and temporal component, these limits could
occur in seperate bands or at different times. The
decision tree identifed bands 5 and 6 from April and
May as giving the best degree of spectral separation
between the cover types. May was identified as the
best single date and February as the worst from the
available imagery. Figure 2.1 shows the coincident
spectral plots for bands 5 and 6 of April and May
with the derived decision boundaries and decision
tree.The sugar beet, beans and peas proved
inseperable. This is because at this time of year
they are all in a pre emergent state, and can
logically be grouped as a bare soil class.
Figure 2.1 Coincident spectral plot and decision tree
for April and May 1980
Beet, 7 = Peas, 8 = Winter Beans, 9 = Coniferous
Woodland, 10 = Deciduous Woodland, ■ = Decision
Boundaries.
Using the band combinations suggested by the decision
tree the following data were classified. Bands 4,5,6
& 7, 18 February 1980, bands 4,5,6 & 7, 12 April
1980, bands 4,5,6 & 7, 18 May 1980, bands 5 & 6, 12
April + 18 May. These classifications would establish
the best single date for crop classification and
determin wether there are improvements from the use
of multitemporal imagery. The classification
algorithm used in each case was the supervised
gaussian maximum likelihood classifier. In each
classification all 10 cover types were included with
0.3% of the data rejected in the thresholding.
Confusion matrices were drawn up by determining the
spectral classes identified for each of the training
sets. It is well established that the analysis of
training data gives optimistic classsification
results, (Swain, 1978, Schowengerdt, 1983) but where
comparisons between classification performance are
required, as in this study, such an approach is
justified. Overall classification accuracy assesments
could not be easily made from the original confusion
matricies as the training sets of individual
informational classes varied in size. Table 2.2 the
confusion matrix from the May data illustrates the
problem.
Table 2.
2 Confusion
matrix
: 18 May 1980
0SR
WW WB
SB
G SBe P WBe WC
WD
U
OSR 151
20
WW
1268 557
17
38
130
WB
53 182
2
11
24
SB
4
73
1 1
1
6
G
4 12
28
3
SBe
6
111 13 14
16
P
13
111 117 20
13
WBe
63 19 38
5
WC
360
1
43
WD
2
279
31
OSR = Oilseed Rape,
WW = Winter Wheat, WB
= Winter
Barley,
SB = Spring
Barley
, G = Grassland,
SBe
=
Sugar Beet, P =
Peas
ì , WBe
= Winter Beans,
WC =
Coniferous Woodland,
, WD =
Deciduous Woodland
Consider the grassland informational class. 28 of
the 47 training pixels were accuratly identified by
the grassland spectral class, however 38 winter wheat
training pixels were incorrectly classified as
grassland. More of the grassland spectral class is
winter wheat than grassland. This only arrises
because of the much larger size of the wheat training
set. The 38 incorrectly classified winter wheat
pixels only account for 1.9% of the wheat
informational class compared with the 28 correctly
identified grassland training pixels which account
for 59.7% of the total.
Improvements can be gained where each class is
weighted on the basis of some known measure of "true"
class proportion. Card, 1982 suggests that "True map
category marginal proportions" can be used to improve
estimates of "proportion correct" for each cover
class. Using such values the output classification
data in a confusion matrix are weighted on the basis
of their anticipated size. Classification accuracies
can then be determined from the weighted matrix in
the usual way. As the classes are no longer given
equal weighting a better estimate of class purity is
given.
The government annual June crop census data can be
used as a weighting factor. The returns from the
county of Suffolk for 1980 were obtained from the
public records office, Kew, Richmond, and used to
weight all the elements of the confusion matricies.
Table 2.3 shows the weighted confusion matrix for the
May classification. Classification accuracy for any
informational class Is calculated by expressing the
diagonal as a proportion of the diagonal plus all non
diagonal elements for that class. Comparisons between
the classification accuracies of the single date and
multitemporal data could then be made.
Table 2.3 Weighted confusion matrix : 18 May 1980
OSR
WW
WB
SB
G
SBe
P
WBe
WC
WD
OSR 2.03
WW
26.0
11.4
0.4
0.8
WB
3.7
12.7
0.1
0.8
SB
0.8
14.1
0.2
0.2
0.2
G
0.4
1.3
4.2
SBe
0.4
7.6
0.9
1.0
P
0.1
0.7
0.8
0.1
WBe
0.04
0.04
0.01
WC
3.0
WD
3.0
(See table 2.2 for key)
Area estimates for each cover type can be made by
using the classification accuracy estimates to bias
correct the pixel counts obtained for each spectral
class.
