cp
"o UNEMUT tte MAC P "$ xk X D
The stratified scenes are subject to unsupervised classification,
leading to 50 significant spectral classes per scene. The result is
compared to reference data, and land-cover types are assigned
(Table 1). Unsupervised classification is given preference over
the supervised technique, because the assignment of classes a
posteriori is generally less time-consuming than the selection of
training areas - especially when working on large data-sets.
This advantage is diminished by the fact that the signatures
generated by clustering will not always meet the expectations
of the analyst, i.e. one class can be represented by several sig-
natures, and two or more classes might fall into one generated
signature. In the first case aggregation of classes will solve the
problem. The second case cannot be solved without additional
information from the image data. Applying supervised training
for the ‘missing’ classes helps to overcome the problem, thus
leading to a hybrid classification with a few signatures defined
over training areas, whereas the majority of signatures results
from the clustering process. Experience gained during the proj-
ect work has shown that the cover types ‘mixed built-up’ and
‘vineyards’ cannot be detected reliably by clustering. There-
fore, these land-cover types are trained applying the supervised
method. The hybrid classification results in a land-cover layer
representing all classes except the alpine ones (Table 1, 14-15).
A different approach was used for the analysis of areas above
the timberline. In these areas the spectral response of land-
cover is influenced by the different illumination angles result-
ing from variation of the terrain. This so called terrain effect
can be significantly reduced by employing ratio images. De-
spite the loss of information caused by image rationing, the use
of the NDVI proved to be sufficient for identification of the
cover types in question. To distinguish between ‘alpine’ and
‘non alpine’ cover types the digital terrain model is intersected
with the classified image. All cover types except forest and
water are defined as alpine areas if located above a certain alti-
tude. For these areas the NDVI is calculated and thresholded
with respect to reference data, thus allowing for the seperation
of alpine vegetation and non vegetated alpine areas.
ciers. Natural vegetation represents all areas with vegetation
cover except forest, which do not result from human activities,
e.g. natural grassland and shrubs. No vegetation stands for
opened spaces with little or no vegetation such as bare soil or
rock. Water comprises water courses as well as water bodies,
wetlands represents non-forested waterlogged areas.
Level I Level II
I artificial surfaces L1 high density urban
L2 low density urban
L3 green urban
L4 industrial/commercial/traffic
L5 mineral extraction sites
IT agricultural areas | II.1 arable land
IL2 vineyards
IL3 pastures
IL4 heterogeneous agricultural areas
III natural areas III.1 forest
IIL.2 natural vegetation
IIL.3 no vegetation
IILA glacier
IV wetlands IV.1 wetlands
V water V.1 water
1 water 9 grass (high)
2 pure built-up 10 grass (low)
3 mixed built-up 11 wetlands
4 pavement 12 shrub
5 gravel 13 forest
6 bare soil 14 alpine vegetation
7 crops 15 non vegetated alpine areas
8 vineyards 16 glaciers
Table 1: example for land-cover classes
3.4 Spatial Classification
The definition of the final land-use types is adapted from the
CORINE land-cover nomenclature (EUR, 1993) and comprises
5 Level I and 15 Level II classes (Table 2). High and low den-
sity urban represent different densities of built-up areas, green
urban areas refers to artificially vegetated areas within an ur-
ban environment, such as parks or cemeteries. Indus-
trial/commercial/traffic areas include industrial structures and
shopping malls as well as large train stations and airports. Min-
eral extraction sites comprises all kinds of surface mines.
These five land-use classes are aggregated under artificial sur-
faces. The second Level I class, agricultural areas, is separated
into arable land, vineyards and pastures. Agricultural areas,
where none of the three land-use types dominates, are defined
as heterogeneous agricultural areas. Forest and natural areas
comprises forest, natural vegetation, no vegetation and gla-
Table 2: land-use nomenclature (adapted from CORINE, 1993)
To derive the land-use classes the spatial postclassification
algorithm is applied to the land-cover layer. As discussed
above, the result of this algorithm strongly depends on the size
of the window which determines the degree of generalisation.
The size of the output pixel is defined by 4x4 input pixels, thus
leading to a cell-size of 100x100m in the final land-use layer.
Based on the experience gained from former studies on spatial
classification (Steinnocher et al. 1993, Ecker et al. 1995) two
different window sizes are defined as local neighbourhood,
depending on the land-use classes. In the first run, artificial
surface classes, forest, glacier, water and wetlands are post-
classified applying a 8x8 pixel window. This size corresponds
to 200x200m, which is suitable for recognising urban struc-
tures without causing too much generalisation. Homogenous
classes such as forest (III.1), glacier (111.4), water (V.1) and
wetlands (IV.1) can be found by examining the portion of the
corresponding cover-type within the window (Table 3: rule 1-
3). To detect heterogeneous land-use types, a composition of
primary classes has to be analysed (rule 4-7), e.g. high density
urban (1.1) is expected to consist more than 7096 of pure built-
up (2) and pavement (4), with pure-built up covering at least
40% (rule 5). All areas which do not meet any condition in the
rule-set are assigned to the rejection class.
rule condition land-use class
(13) > 50 % TII
f(17) > 50 % 111.4
f(1) >50% V.1
f(9) > 50% :TV.1
(£2) + f(4)} > 70 % AND f(2)> 40 % :L1
Q2) + f(4)} > 70 % AND f(4) > 40 % :14
{f(4) + f(5)} > 70 % AND f(5) > 40 % :L5
{f() + f(4)} > 50 % AND f(3)> 30 % :I2
ELSE :0 (rejected)
O0 —-1 QV tA SR UN —
with f(n): frequency of land-cover type n (class numbers refer to tables 1 and 2)
Table 3: example of rule-set I
The rejected areas are used as a mask for the primary classifi-
cation layer, thus leaving only non-postclassified areas for the
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
