vegetation. This non-patch typically meets the criteria
for a matrix, because it covers most of the area and
constitutes a single area. Moreover the omnipresence of
the herbal vegetation determines the conditions for the
germination and growth of shrubs and as such controls
landscape development. Obviously, the patches are
best represented by objects, while the continuous
character of the matrix should be modelled as a field.
object boundary resolution
Fig. 2. The concept of fields in objects. One of the three
different object types (left) retains its field character
(right).
In the previous section it appeared that the modelling
of fields in objects is a two step process. in the first
phase the discrete terrain features are represented by
objects (fig. 2left). In the second phase the objects
covering continuous terrain features have to recover
their field character. This can be achieved through the
definition of a field within the object (fig. 2right).
Obviously, internally homogeneous objects remain
discrete. Because of the construction process, where a
field is created within a heterogeneous object, the
concept is called ‘fields in objects’. Indeed this concept
enables the modeller to represent the notion of objects
located in a field, like the distribution of solitary shrubs
over a continuously varying herbaceous vegetation.
In the next sections we will elucidate the
construction of fields and objects with digital
photogrammetry.
3 IMAGE PROCESSING AND INTERPRETATION
3.1 Production of digital orthophoto mosaics
A typical list of processing steps in the transformation of
analogue photographs to digital orthophoto mosaics is
scanning, radiometric correction, geometric correction
and mosaicing. It is no problem to obtain pixel sizes and
positional accuracies better than 1 metre. The
radiometric correction is discussed in more detail.
Density variations in a photograph are not
solely related to variations in terrain conditions. Factors
influencing the density variation having nothing to do
with the actual terrain characteristics are termed
extraneous effects (Lillesand and Kiefer, 1987),
Extraneous effects are of two general types: geometric
and atmospheric. The magnitude of geometric factors
varies structurally over the image while atmospheric
effects are constant throughout the image. Obviously,
these effects prevent false colour photographs from an
accurate quantitative interpretation.
Clevers and Van Stokkom (1992) present a
method to undo false colour photographs from all
extraneous effects in order to derive reflectance factors,
However, the method requires some information on
camera and film characteristics and reference
measurements in the field, which are usually not
available. For classification purposes relative differences
in density, which can be attributed to differences in the
terrain, suffice. The latter is achieved by only removing
geometric deviations in density. Lilesand and Kiefer
(1987) enumerate some important geometric effects
influencing film density:
- light fall-off caused by a geometricaily based
decrease in illumination at the film plane with
increasing distance from the centre of the
photograph.
- differential scattering by the atmosphere.
- non-lambertian reflection by natural objects.
- differential shading caused by relief in the
vegetation cover, especially shrubs and trees.
Contrarily to the first two effects which are indifferent to
terrain cover, the latter two factors are dependent on
the terrain surface characteristics, e.g. the deviations are
higher for woody vegetation than herbaceous
vegetation. We applied a two dimensional second order
polynomial function to correct for these deviations. The
parameters are obtained by regressing the function
through an extensively sample set obtained in
herbaceous vegetation.
3.2 Crisp and fuzzy image classification
Radiometrically corrected orthophotos are ready for
digital image interpretation. A straight way of
interpreting digital images is the regression of spectral
data with quantitative variables measured in the field
like biomass and vegetation cover. Because the spectral
patterns obtained from natural scenes are usually very
complex, classification seems a more robust
interpretation technique. Unfortunately remotely
sensed data have tended to be crisply classified
regardless of whether the vegetation exists as a well
defined mosaic or as a series of continua (Wood and
Foody, 1989). Consequently, many classification errors
can be attributed to artificial boundaries in an image
where in reality gradients exist. Crisp classification has
to be treated with some caution in patterns of natural
216
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
lan
cor
clas
que
rep
bec
vali
exc
cla:
19€
The
hiel
fror
der
nor
COV
top
isol
isol
inte
sub
her
sub
gra
sub
Giv
apr
typ
wh
Cor
res
and
the
dea
val
Whe
me
inst
COV
spe
app
Sub
inte
fuz:
veg
(hs,
pro
ah
hav
Clas
qua
rese
mer
