second processing step. The remaining land-use classes, agri-
cultural and natural areas, are postclassified with a window
size of 16x16 pixels, which equals 400x400m on the ground.
The higher degree of generalisation corresponds with the
definition of these classes, as they are to represent dominant
forms of land-use. This is considered in rule-set II (Table 4)
which first examines the dominant occurrence of agriculture
(rule 1-3) and natural vegetation (rule 4). If no dominant cover-
type is found, a composition of cover-types is considered (rule
5), Le. any combination of different agricultural cover types or
of agriculture with natural vegetation covering more than 80 %
will be classified as heterogeneous agricultural areas (11.4).
Rule 6 and 7 test for the dominant occurrence of vegetation
within alpine areas.
rule condition land-use class
1 {f(6) + f(7)} > 70 % AL]
2 f(8) > 70% 112
3 {f(10) + f(11)} > 70 % 1.3
4 {f(12) + f(13)} > 70 % “IL 2
S {f(6) + f(7) + f(8) + f(10) +
+ f12y+ 13 >80% 114
6 {(13 + 14)} > 50 % :HF2
7 {(15 + 16)} > 50 % 11.3
ELSE :0 (rejected)
with f(n): frequency of land-cover type n (class numbers refer to tables 1 and 2)
Table 4: example of rule-set II
The two resulting layers are then intersected, giving priority to
the less generalised structures of the first postclassification.
Assuming an optimum design of the rule-sets, all pixels are
assigned a land-use class at this stage. Experience has shown,
though, that up to 596 of the pixels will fall into the final rejec-
tion class, i.e. they are rejected by both rule-sets. The majority
of these pixels occur as single pixels within a classified neigh-
bourhood, e.g. on the border between two land-cover types,
where neither reaches the majority within the local neighbour-
hood, though both cover types are close to it. These pixels can
easily be classified by assigning the relative majority of a 3x3
neighbourhood.
In addition to these *border cases', larger rejected areas might
occur, resulting from particular combinations of cover types
not considered in the rule-sets. For classification of these areas
we propose interactive post-editing rather than setting up addi-
tional rules, as the consideration of all possible combinations of
cover types seems unrealistic. Post-editing might also be neces-
sary for some of the artificial surface classes, where the defini-
tion of the class is based on the spatial context rather than on
the local pattern of cover types, e.g. green urban areas are
defined as vegetation within an urban environment. Similar
problems might occur in the separation of industrial areas from
particular mineral extraction sites such as gravel-pits. Being
aware of these problems the post-editing process can be con-
centrated on the doubtful areas, thus reducing the time needed
for interactive work to a negligible amount within the entire
project.
4. RESULTS AND DISCUSSION
To assess the quality of the land-use model it was imported to a
GIS and compared to the CORINE land-cover data-base. Six
map sheets were chosen from available ‚parts of the CORINE
data-base, each covering about 500 km“. Therefore, the total
844
number of pixels evaluated comes up to about 300.000, which
equals 3.5 ?6 of the entire area investigated. When selecting the
map sheets, attention was paid to consider different forms of
the Austrian landscape. The test sites contain all classes except
glacier as no layer containing glacier is available yet.
For comparison the CORINE vector layers are converted to a
raster representation, using the raster of the land-use model as
geometric reference. In addition, the 44 CORINE classes are
aggregated according to the 15 land-use classes. CORINE de-
fines the smallest mapping unit with 25 hectares, therefore all
areas smaller than 25 hectares are eliminated in the land-use
model and redefined applying an iterative majority filter.
Intersection of the two models at all test sites allows for the
derivation of confusion matrices. Table 5 shows the overall
confusion matrix for Level-I and Level-II land-use classes,
including totals and percentages of identical results. The rows
of the matrix represent the CORINE land-cover, the columns
are the result of the automated classification. With a sample of
less than 100 pixels, Mineral extraction sites (1.4) is underre-
presented and therefore not considered in the matrix. The fol-
lowing discussion is based on comparison of conflicting areas
to reference data. It concentrates on the examination of sys-
tematic errors of both the automated and the visual classifica-
tion.
When analysing the confusion values of Level-I classes a high
correspondence between the two data sets can be observed (>90
%). Some deviations are found between artificial surfaces (I)
and agricultural areas (IL), resulting largely from differences in
low density urban (1.2), and between agricultural areas (II) and
natural areas (III), which is due to a different classification of
forest.
Within agricultural areas there is a significant confusion be-
tween heterogeneous agricultural areas (11.2) and arable land
(IL.1) as well as pastures (11.3). The reason for this disagree-
ment becomes obvious when comparing maps of both models
(Figure 3). Whereas the automated classification reacts rather
sensibly to local variations of land-use, visual interpretors have
a tendency to integrate larger areas within one class (compare
patchy pattern in the left upper part of the left map with same
area in the right map in Figure 2).
The most critical confusions are found within natural areas
(III). Forest (111.1) was slightly overestimated in the automated
approach, which results in differences of totals for this class
(compare sum of column III.1 with sum of row IIL1 in Table
4). Comparison with reference data has shown that long and
narrow valleys which are surrounded by forest, get lost during
the postclassification process, although they are larger than 25
hectares and therefore classified in the visual interpretation.
These patterns are typical for the alpine landscape and cause
the majority of confusion between forest and agricultural areas.
Natural vegetation (111.2), which occurs predominantly in the
high alpine regions, is strongly confused with forest (111.1) and
no vegetation (II1.3). Detailed analysis has shown that areas
clearly recognised as forest in the reference data were actually
classified as natural vegetation in the CORINE data-base. This
may be partially explained by the fact that dwarf-pines, classi-
fied as forest in the automated process, were interpreted as
natural vegetation in the CORINE land-cover maps. Though
dwarf-pines cover large areas in the Alps, it is doubtful that
they are the only explanation for this confusion. We assume
that the different illumination angles are the essential reason for
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
