bul 2004
predicted
ss of the
with the
he model
he model
with the
rrors, the
an fitted
| species
d the G-
(G) is a
onse and
night be,
using the
[2]
oZ xi) is
neasured
'ediction,
ent, and
; reliable
der et al.
ss of the
iodel had
for each
iting the
window
f values
cel of the
'erations.
| lichens,
using the
'jents) as
are maps
el in the
ation of
1 in table
0.59 for
e power,
een 0.48
ained. In
sampling
sampling
variables
ly seven
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Models Lichen Trees Rocks Soil
total
Goodness of ft 0.68 0.79 0.61 0.58
(r model)
Brdkmepowr — x 0.79 054 0.48
(r reference)
MAE (model) 7.84 6.65 8.77 1.01
9376 Quante of... 15.21:3 15:820/25:20.1 10450: 3.87
error (model)
Bias (model) +1.09 +152 +3.26 +0.46
G 0.52 0.63 0.37 0.40
Table 2. Means of 100 runs for validation of the four models of
the species richness for all lichens, on trees, rocks soil
Fig. 5 shows the maps of the predicted number of species for all
lichens in LUUI (low intensively use) and LUU6 (high
intensively use). Areas with low numbers of species are mapped
gray whereas areas with high numbers of species are white.
Prediction for LUUI
TN M S
0 7 #0 300 Meters GWSL, 2002 9 75 150
LL id
Orthoimage LUU6
b uere e
E M oH . n
300 Meters Swit,
Prediction for LUUG
MO Meters
Figure 5. Maps of predicted species richness for all lichens for
LUUI and LUU6 with their corresponding CIR orthoimages
Our approach in distinguishing between 1st and 2nd level
explanatory variables allowed us to assess their contribution to
the corresponding model. This was an important step for the
development of the final models and helped us to drop the
variables that contribute less to the model. Particularly the use
of simple spectral and textural information values of the CIR
orthoimages which is linked to spectral reflection and spatial
heterogeneity of the vegetation cover, respectively, produced
best results. The implementation of additional 2nd level
explanatory variables in fact improved model accuracy again —
with the exception for lichens on trees. For this model best
accuracy (r model of 0.79 and G 0.63) was produced with the
single use of variance nir and its quadratic term whereas the
implementation of additional explanatory variables slightly
deteriorated the model's accuracy. In this case the number of
849
species is directly related to a high heterogeneous vegetation
cover such as forest borders and forest itself. The nine land
cover types extracted for this study are based on what was
supposed to be detectable in CIR orthoimages, and what was
regarded to be of importance for the lichen diversity. The main
advantage of the application of an object-oriented image
classification method is that it allowed us to define land cover
types according to the needs of lichen experts. Thus land cover
classification applied in this study in combination with image
segmentation methods was an important step in the
development of the models. The main disadvantage was the
relatively high complexity and required amount of time of
object-oriented image classification methods.
4. CONCLUSION
This study reveals that the application of homogenous and
reproducible land cover information derived from remotely
sensed data as basis for the model is adequate. The accuracies (r
reference) obtained for both model lichens-on trees (0.79) and
for all lichens (0.58) can be regarded as good for the application
purposes by lichenologists.
The crucial question is how we can improve our models for
lichen species richness? In this study we were confronted with
several problems concerning ecological modelling. According
to Leathwick et al. (1996) a model used for biodiversity
assessment should also be general, which means applicable in
other regions or different times. Furthermore, according to
Fielding and Bell (1997) the lack of validation and uncertainty
assessment of models remains a serious issue in ecological
modelling. Finally, according to Austin and Gaywood (1994) a
model used for biodiversity assessment should not only be
precise but also be ecologically sensible, meaningful and
interpretable. Meeting all the suggested requirements turns out
to be nearly impossible in our case. E.g. the particular model
developed here has been applied only for six test sites. Thus,
the resulting variables of the presented linear models may be
used for calculating species richness in neighboring regions of
the Entlebuch with similar vegetation cover and landscape
structures. Applying the model to other regions is a well-known
problem (Iverson and Prasad 1998).
There are four points to remember: First, linear regression
models can be used to predict lichen diversity, but strongly
depend on the sampling design of the lichen relevés. Thus the
distribution of the lichen data should be analyzed further.
Second, possible hotspots were calculated and may help in
reducing field surveys and could be useful for possible
conservation efforts. The resulting explanatory variables of the
presented linear models may be used for calculating species
richness in neighboring regions with similar landscape
structures. Third, we can confirm that the application of
homogenous and reproducible land cover information derived
from high resolution remote sensing data as basis for the model
is very adequate. This means that not so well-known areas can
still serve as a basis for building the methods. Fourth,
explanatory variables can be rapidly derived from high
resolution remote sensing data and distinguishing between 1st
and 2nd level of detail proved to be a good method for the
development of the models.
This method cannot replace lichen surveys altogether, but it can
be used to target focused lichen forays in the future. Finally, it
should be noted that this method cannot produce any
information on lichen species abundance, dynamics, or
viabilities; it only indicates the potential presence or absence of
species. .
