bul 2004
f1kmx
(see fig.
rest and
id of the
epresents
ording to
entage of
ata set is
data are
ls. In our
stal of 96
of 200 m
set up by
-/- 0.5 m.
on the 96
(56.41 m
ites were
| the 16
dius)
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
0 (3 . 420 Meters NE 20
LLL ls 1
Figure 3. Lichen relevés were carried out at 1-12 randomly
selected collecting sites
At each of the 12 collecting sites, lichen relevés were carried
out on three different substrates, i.e. trees, rocks and soil -
representing all major lichen substrates which could be affected
by changes of the agricultural and forestry management.
For relevés on trees the nearest tree within the border of the
sampling plot was selected and for relevés on rocks, the nearest
saxicolous object within the border of the sampling plot was
selected (for both starting from the center of a collecting site).
For relevés on soil in the center of each collecting site a
frequency grid of 50 x 40 cm mesh size 10 cm) was placed on
the ground. For each lichen species the number of unit areas (10
x 10 em) where the species occurred was counted (a value
ranging from 1 to 20. Since delimitation of individuals is often
difficult or even not possible in lichens, we used the number of
occupied unit areas as abundance measure.
As the calibration data set every second sampling plot was
chosen. The remaining 48 sampling plots served as reference
data set.
2.2.2 Calibration data: In order to calibrate our model of
prediction of species richness we tried to find biological /
ecological meaningful features as explanatory variables. For
this purpose we used original and derived spectral and spatial
information of airborne remote sensing data.
Six digital CIR orthoimages of the years 1999 and 2001 served
as the basis for this study. Each orthoimage covers an area of
approx. 2 square kilometers. The scale of 1:10'000 provides a
ground resolution of 0.3 m. Each image offers three color bands
of numerical information with 256 intensity levels: visible
green (500-600 nm), visible red (600-700 nm) and near infrared
(750-1000 nm). Additionally to the original spectral and spatial
information several derivatives of the CIR orthoimages were
calculated. For our approach we decided to extract derivatives
both using standard methods and additional expert knowledge.
Furthermore we used a digital terrain model with a spatial
resolution of 25 m (DHM25 O 2003 Bundesamt für
Landestopographie, DV 455.2) and digital surface models
(DSM). A spatial resolution of 0.5 m was chosen for all data
sets used in this study.
To assess and categorize the contribution of ecological
meaningful variables to the model we decided to distinguish
between two levels of detail. Ist level variables provide
information of. spatial heterogeneity, spectral reflection,
absorption and transmission, chlorophyll content and above-
ground phytomass of vegetation cover. This implies simple
image processing methods (standard methods) of the CIR
orthoimages, and was performed without additional expert
, 847
knowledge, e.g.
channels
(red, green, NIR)
of biologists. In addition to the three original
several new variables were
generated using both spatial and spectral information within a
moving window of different sizes. The wider the window, the
more these new variables tend to reflect features of the
landscape. The window size of 6x6 pixels turned out to be the
most adequate. Table 1 lists all variables applied in this study.
ID Name Comments
1" level variables
Mean, majority,
minority, sum of:
1-3 Red, green, NIR original channels of CIR
orthoimage
4 Ratiol Channel green / Channel (red +
NIR)
5 Ratio2 Channel red / Channel (green +
NIR)
6 Ratio3 Channel NIR / Channel (red +
green)
7-9 Variance red, returns variance in a moving
green, NIR window
10-12 Skewness returns skewness in a moving
window
13-15 Contrast red, returns contrast in a moving
green, NIR window
16 Vegetation Index NIR - red
17 NDVI NIR - red / NIR + red
2" level
variables
18-20 Fraction of land forest, non-forest, non-vegetation
cover (3 classes)
21-29 Fraction of land forest, grassland light, grassland
cover (9 classes) dark, rock&gravel&soil, sealed
surface, single trees & hedges,
shadows, wetlands, water bodies
Table 1. A total of 29 explanatory variables were derived
On the 2nd level, new variables based on Ist level variables
were built using expert knowledge and field experiences. To
meet these requirements, new image processing techniques
were applied to produce homogenous objects and well defined
object edges. Two land cover classifications were performed: 1)
a simple classification only distinguishing between forest, non-
forest and non-vegetation and 2) a more detailed classification
distinguishing nine land cover classes, representing the three
lichen substrates of the field survey: 1. forest, 2. grassland light
(mown and intensively used), 3. grassland dark (unmown and
not intensively used), 4. rock & gravel & bare soil, 5. sealed
surface, 6. single trees & hedges, 7. shadows, 8. wetlands and 9.
water bodies. For this classification an object-oriented approach
was applied. The method is based on hierarchical segmentation
not only of the CIR orthoimages but also of their derivatives
(Baatz and Schápe 1999).
To summarize, we produced a total of 29 explanatory variables
for the model. 17 of them were allocated to 1st level variables,
mainly based on simple reflection values of the three channels
of the CIR orthoimages as well as on spatial information. The
remaining 12 were allocated to the 2nd level variables.
Finally, in accordance with the lichen relevés that are
representative for a 56 m circle, for each variable the sum of
values was calculated within a 56 m radius circle for each of the
96 sampling plots. This was performed using a moving window
approach - in our case a moving circle (see fig. 4).
