be found
different
ıpervised
| before
ormation
different
matrices
> satellite
Nis image
accuracy:
esolution
using an
-database '
data was
y data for
vhich are
wer than
e training
rd)
ikelihood
4,
standard.
total
242
519
256
440
1457
standard
accuracy
b
b
o
Jo
yout 70%,
'W classes
data set.
Classification accuracy of the training set turned out be
almost 90%. Both producers and users accuracy of
irrigated tree crops are good, 83.9% and 99.6%
respectively which is probably due to the more
homogeneous cover of these fields in comparison with
others. Moreover, irrigated tree crops are mainly confined
to the relatively large flat areas near the river.
Producers and users accuracy of the other three land
cover types range between 55 and 70%.
Misclassifications between these cover types occur
mutually for which the status of the land cover types in
May 1995 may give an explanation. In this rather dry
year, the other cover types had all a low vegetative cover
which generally affected their reflectance patterns. Most
annual crops were in a bad condition or died, so that
reflectance was for a considerable part from almost bare
soil. Rain-fed tree crops in the area were widely spaced.
Therefore the sensor mainly detects bare soil or dead
undergrowth. Often the semi-natural vegetation was not
fully developed and without its normal green color.
Approach 2 Topographic normalization before
supervised classification
Table - 3 Classification accuracy report after
topographic | normalization with the Lambertian
reflectance model
Cover class producers accuracy | users accuracy
irr. tree crops 81.2% 99.6%
rain-fed tree crops 78.9% 69.4%
semi-nat. vegetation | 48.2% 65.3%
annual crops 76.6% 67.0%
Overall accuracy 73.0%
The classification accuracy of the normalized image
obtained by applying transformation based on Lambertian
reflectance model shows only about 3% increase in
overall accuracy due to normalization (table 3). The
accuracy for rain-fed tree crops increased, but for annual
crops decreased, while for semi-natural vegetation the
producers accuracy decreased and the users accuracy
increased. Possible reasons for the only small increase in
accuracy may be: (1) mitigation of the specificity of a
cover and slope combination may cause misclassification
(2) the Digital Elevation Model (DEM) might have
introduced some errors in slope and aspect values (3)
although the spatial resolution of the DEM and the
satellite image are comparable, there is no perfect
geometric match (4) Lambertian reflectance assumes only
direct irradiation and ideal cover types, reflecting the
same in all directions. This is probably not the case.
The increase in accuracy for rain-fed tree crops may be
due to the fact that rain-fed tree crops grow on many
different slopes, while the other cover types grow
preferentially on only a limited number of slopes: semi-
natural vegetation mainly on steep slopes, annual crops in
rolling areas, and irrigated tree crop in flat areas.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
The overall accuracy using a non-Lambertian reflectance
model is comparable (2% less) to the approach using
Lambertian reflectance, as opposed to the expectation.
This may be partly due to the fact that the Minnaert
constant which is needed in the calculations, is only
based on one crop type. This is illustrated by the lower
accuracy of annual crops, caused by the deviating
structure from the selected cover type.
Table 4 Classification accuracy report after topographic
normalization with the non-Lambertian reflectance model
Cover class producers accuracy users accuracy
irr. tree crop 83.396 99.6%
rain-fed tree crop 78.8% 67.9%
semi-nat vegetation | 47.1% 69.3%
annual crops 71.7% 62.8%
Overall accuracy 71.7%
Approach 3 Contextual information to
supervised classification.
improve
In the preceding part methods were described to
compensate for the effect of topography. In this part use
is made of relations between topography, soil and land
cover. The rules are applied after the supervised
classification.
Two methods were examined to obtain decision rules:
e Expert knowledge of a soil scientist,
e Discriminant analysis (statistical analysis of ancillary
data)
The classification accuracy report using expert knowledge
shows a considerable increase in accuracy for almost all
cover types (table 5). Especially the producers and users
accuracy of rain-fed tree crop and annual crops are
higher, while the users accuracy of semi-natural
vegetation and the producers accuracy of irrigated tree
crops increased. The users accuracy of irrigated tree crops
shows a slight decrease while the producers accuracy of
semi-natural vegetation shows a slight increase. Overall
classification accuracy is increased by about 18% by this
method. In figure 5 the classified image is shown.
Table 5 Classification accuracy report using expert
knowledge
Cover class producers accuracy | users accuracy
irr. tree crop 98.9% 98.9%
rain-fed tree crop 95.3% 82.8%
semi-nat vegetation | 59.3% 91.6%
annual crops 96.9% 85.4%
Overall accuracy 88.6 %
The producers accuracy of irrigated tree crops shows
about 10% increase mainly due to misclassified rain-fed
tree crops, which are reclassified as irrigated tree crops
based on a decision rule. The users accuracy of irrigated
tree crops is not increased since in the study area some
347
