presented in figure 3 is based on landform element
information.
Soil types, studied at facet and site level, become an
aggregated attribute at higher levels; it is called a soil
association.
Often the higher level is very predictive as to a particular
soil property: the Hills in our example consist mainly of
marls (parent material) on which dense clayey soils
(figure 3) develop unsuitable for tree crops. However,
when studying the landform element composition in more
detail it turns out that certain hills are capped with fluvial
remnants where olives grow well while others consist of
steep sandstone ridges with shallow soils, where only
natural vegetation will do. The lower level information in
this example enables us to adjust general statements based
on higher level information.
The rules to predict soil properties have therefore a nested
format just as the terrain units of the land system
hierarchy (figure 2) on which they are based. First a
general rule is formulated based on the major landform.
That rule is then conditioned by a nested subset of rules in
accordance with these lower level nested terrain units as
follows:
1. If major landform is Hills consisting mainly of marl
then we expect swelling clays
1.1 But if the land form element within the Hills is a
fluvial terrace remnant and/or erosion glacis, then
we expect permeable medium to coarse textured
soils
1.1.1 But if slopes of the terrace remnant are steeper
than 30 % then soils are very shallow and if the
unit corresponds with the lower slopes of the
erosion glacis then soils are dense clayey.
Because soil properties predict land cover performance
experts can use the same nested rules to predict land
cover occurrence starting from major land form level and
adjusting for specific conditions at more detailed level.
1. If major landform is Hills consisting mainly of marl
then we expect cereals
1.1 But if the land form element within the Hills is a
fluvial terrace remnant and/or erosion glacis, then we
expect olives
1.1.1 But if slopes of the terrace remnants are steeper
than 30 % then we expect semi-natural vegetation
and if the unit corresponds with the lower slopes
of the erosion Glacis then we expect cereals.
These type of rules have been applied to improve the
classification results. Some other typical rules are:
e In serpentinite ultramafic rocks (MuT, figure 3) only
semi-natural vegetation is present.
e On hills in soft rocks with marl-rich parent materials
(HsZ) no rain-fed tree crops are found.
e On hills in hard rocks from Gneiss (MmG), semi-
natural vegetation and rain-fed tree crops may occur.
Annual crops and irrigated tree crop can not be found
in this area.
5. RESULTS AND DISCUSSION
In this section the results are presented of the 3 different
approaches for classification: (1) standard supervised
classification (2) topographic normalization before
supervised classification (3) use of contextual information
after supervised classification. The results of the different
approaches will be compared by evaluating error matrices
and classification accuracy. In all cases the same satellite
image was used to extract land cover classes. This image
was georeferenced with acceptable spatial accuracy:
within the total RMS error of 1.1 pixel (resolution
30m*30m). Classification accuracy was assessed using an
independent data set. Land cover data of the geo-database
of Alora were used. The collection of these data was
intended to record cover types and not as training data for
classification, so it even may include fields which are
having mixed pixels. Accuracy is therefore lower than
what would be achieved using the accuracy of the training
set.
Approach 1. Supervised classification (standard)
The result of the standard supervised maximum likelihood
classification for the 4 classes is given in figure 4.
Table-1 Classification error matrix for standard.
supervised classification.
classified | reference data
data
irr. tree | rainfed | semi- annual total
crop tree natural | crops
crop veget.
ir. tree | 241 0 l 0 242
crop
rain-fed 36 348 45 90 519
tree crop
semi-nat. 6 63 151 36 256
veget.
annual 4 84 77 275 440
crops
287 495 274 401 1457
Table - 2 Classification accuracy. report for standard
supervised classification.
Cover class producers accuracy | users accuracy
irr. tree crops 83.9% 99.6%
rain-fed tree crops 70.3% 67.0%
semi-nat. vegetation | 55.1% 58.9%
annual crops 68.6% 62.5%
Overall accuracy 69.7 %
Table 2 shows that the overall accuracy is about 70%,
which is reasonable taking into account the few classes
and the evaluation with a fully independent data set.
346 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
Clas
almc
irri g
resp:
hom
othe
to th
Prod
cove
Misc
mutt
May
year,
whic
annu
refle
soil.
Ther
unde
fully
App
supe
Tabl
topo,
_refle
| Cove
irr. tr
rain-f
semi-
annu;
Over
The
obtai
refle
over:
accu
Crop:
prod
incre
accu
cove
(2) |
intro
altho
satel]
geon
direc
same
The
due
diffe:
prefe
natur
rollir
