semi-natural vegetation also occurs in flat areas along the
river which was not considered in the decision rules.
Table 5 also shows that the decision rules were not
discriminate enough to separate annual crops and semi-
natural vegetation precisely. Misclassification occurred in
the major landscape unit HsZ (Fig. 3), where mainly
annual crops occur, but where in areas with steep slopes
semi-natural vegetation occurs. Although we tried to cope
with this by applying rules based on slope steepness, the
geometric accuracy of the DEM was not enough to
achieve a higher accuracy.
Alternatively it was tried statistically with discriminant
analysis to obtain rules for classifying cases into one of
the land cover classes from the geo-data set. Agronomic
data were compared with elevation, slope and aspect
derived from the DEM. A case is classified, based on its
discriminant score, in the group where the posterior
probability is the largest. In table 6 the results are shown.
Table - 6 Classification results of discriminant analysis
Actual Predicted group membership
| group
no irr. tree | rainfed | semi- annual
of crop tree natural | crops
cases crop veget.
irr. tree 242 234 6 1 ]
crops 96.7% 2.505 0.4% 0.4%
rain-fed tree 78 28 39 1 10
crops 35.9% | 50.0% 1.3% | 12.8%
semi-nat. 21 6 14 1 0
vegetation 28.6% | 66.7% 4.8% 0%
annual 61 28 23 0 10
crops 45.9% | 37.7% 0% | 16.4%
Percent of “grouped “ cases correctly classified : 70.6%.
Since this approach gives an overall classification
accuracy of only about 71%, it was decided not to make
rules based on this analysis. Therefore in this study we
only use rules based on expert knowledge, which gives a
considerable increase in accuracy.
6. GENERAL DISCUSSION AND CONCLUSIONS
In this study a number of methods are examined to
improve classification accuracy of remote sensing images.
The approach of topographic normalization, using
Lambertian or non-Lambertian reflectance model, does
not give a significant difference in accuracy for the study
area. Land cover classes show a strong topographic
specialization. For example, the land cover irrigated tree
crops grows on flat areas near the river, semi-natural
vegetation grows at higher elevation and on steep slopes
etc. This approach, removal of ‘topographic information’,
will probably be more successful in areas where
topographic specialization is less strong.
The second approach using soil and topographic
information in classification gives an 18% increase in
348
overall accuracy of the image for the study area. The
different land cover types in the study area each have
their specific environment, so that knowledge rules are
effective. This implies that in comparable areas expert
rules will improve the classification accuracy
considerably. However the method demands a good
understanding of the relation between terrain units and
land cover and of the significance of the levels of the land
system hierarchy in order to formulate the rules. When
that is the case rules based on statistical analysis will not
give better results than expert knowledge at
corresponding aggregation levels.
The extent to which the added data and knowledge may
be expected to alter the classification accuracy of a single
object class, depends also on the spectral discrimination.
When spectral class discrimination is poor, the extra
information based on contextual knowledge can improve
the classification.
Expert rules should always be formulated explicitly, at the
right aggregation level. Rules should be nested in case
they combine knowledge from different aggregation
levels. If an expert rule is too rigid (for example in hilly
areas no irrigated tree crops occur) then a monitoring
system may overlook spots where irrigated tree crops do
occur. General rules may be of help in a good spatial
generalization of the area.
Further to this, we expect that in many areas also spatial
expert rules, for instance describing crop as a function of
distance to rivers, may improve the accuracy of the
analysis.
The following general conclusions can be drawn:
|l. Topographic normalization, both assuming
Lambertian or non-Lambertian reflectance, does not
show significant effect on land cover classification
accuracy of a rcinote sensing image of the study area.
The main reason for this seems to be the topographic
specialization of the different cover types.
2. Statistical analysis of land cover data in combination
with topographic data derived from a DEM does not
show an improvement in classification. accuracy.
Probably this is caused by the resolution of the DEM,
which seems still too small.
3. The accuracy of land cover classification in the study
area improves considerably, when knowledge based
on landscape and soil data and its relation to land
cover is included. An average improvement of 18% in
land cover classification is achieved.
4. The measure of improvement in the classification
accuracy as a result of the application of knowledge-
based rules depends on:
e the degree of generalization in defining land
cover classes
e The level of aggregation at which rules are
applied and the degree of adjustment of these
rules with more specific information from lower
levels
e the resolution of the DEM and the thematic
resolution and accuracy of field data.
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
Col
terr.
sen:
Fre:
low
trait
Hut
Lan
imp
rem
Jans
base
ima,
No
Kon
and
inte;
anal
Intei
247
Liya
data
remc
rural
Mol
topo
Infoi
Rich
anal
Schu
land
infor
MSc
Univ
Wiel
HP.
soils
conte
Chac
cong
Geol
Gran
The :
a larg
imag,
