— 0 v tt.0€0 Ww
Do nn SS
a O
slave). The colour infrared aerial photos have been geometrical
corrected by registering them to the German base map (scale 1:
5000). By stitching the results together, the single scenes and
images could be combined to a mosaic.
6. CLASSIFICATION AND RESULTS
Classification is the process of sorting pixels into a finite
number of classes, or categories of data, based on their values.
If a pixel satisfies a certain set of criteria, then the pixel is
assigned to the class that corresponds to that criteria. The
result 1s a (GIS) file whose values represent known thematic
categories such as landcover or vegetation types. For the first
part of the classification process, the computer system must be
trained (supervised classification) to recognise patterns in the
data. Training is the process of defining the criteria by which
patterns are recognised. The result of the training process is a
set of signatures (area of interest, abbreviated as AOI), which
are statistical criteria for a set of proposed classes.
We used the maximum likelihood algorithm as the method for
a supervised classification process. Figure 4 gives an overview
of an ideal integrated GIS and remote sensing environment. It
shows how various data sets can be integrated as layers in a
GIS to start monitoring analyses over time.
A hierarchical method is used for extracting bogland classes
with respect to the environmental protection goals. A highly
accurate classification of the following classes was
accomplished: deciduous- and mixed forest, coniferous forest,
water, very wet areas, meadowland/farmland with vegetation,
meadowland/farmland with partly vegetation, meadowland/
farmland without vegetation, peat quarrying with maximum of
50% vegetation, de- and regeneration stages. Statistical
accuracy analysis shows an accuracy value of over 90% for the
classification (accuracy assessment method with independent
random points). An extended approach for fine tuning of 15
bogland classes gave an accuracy of about 79%. This approach
assigned single bogland areas with good results. For a state-
wide approach the first method is preferable with respect to the
accuracy. An overview of the hierarchical method is shown in
figure 3.
From the ground truth campaign and the classification results,
the following lessons for an operational monitoring concept
could be learned:
* Time variability differs for different areas (e.g., between
forest and meadowland); therefore, different strategies
have to be developed for each class;
e For a prolonged period monitoring concept, it makes sense
to generate individual mask layers (for example with
farmland, meadowland, forest, urban areas, peatland, de-/
regeneration stages, and water);
e These mask layers could be generated from a classification
or vector data-sets with different data sources;
* A meadowland class could be generated or extracted by
using the Normalised Difference Vegetation Index (NDVI)
in several April scenes. Those areas which always show a
high NDVI should indicate meadowland, whereas, due to
crop rotation, arable land will be vegetated only in a
limited number of years [Reinke, 1995].
1st rough classification
~~ Wi
<. mas >
i
de-/regeneration i
peatland stages agricultural areas
3 4 E ER
< classify
SE
TEE
classify
I
Pe EN
lassify >
s em
SC
y
ac
a
A
deciduous- and coniferous forest ater
mixed forest and calluna Wi
- dense birch area
- cotton grass/caluna
- peat quarrying by
milling machine
- peat quarrying by
excavator or cutting - cut meadowland
- farmiand without - de-/regeneration stages
vegetation >= 50% dense areas
- de- /regeneration - mixed de- /regeneration
stages stages
- cereals / meadowland
with vegetation
- cotton grass /birch (596)| |- de- /regeneration
stages with molina
machine - meadowland with caerulia, myrica gale
- peat quarrying 2596 vegetation (25%)
vegetation - meadowland with partial| |- corn
- peat quarrying >= 50% vegetation - meadowland without
vegetation - molina caerulia vegetation
fine classification with
15 classes d
Figure 3: Hierarchical Classification Method
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
