1987).
metric
factors
spheric
ously,
om an
sent a
om all
actors,
lon on
erence
ly not
rences
in the
noving
Kiefer
effects
based
e with
of the
cts.
in the
rees.
rentto
ent on
ns are
3CEOUS
| order
1s. The
nction
ed in
dy for
ray of
;ectral
e field
pectral
ly very
robust
motely
ssified
a well
d and
errors
image
on has
natural
landscapes. Crisp classification is particularly suited to
construct objects of different object classes.
In natural landscapes a continuous type of
classification has to be applied to accommodate the
quantification of ecological gradients. À proper
representation of gradients is of ecological importance,
because these situations often posses high natural
values. The notion of continuous classes can be
expressed with the mathematical concept of fuzzy
classification (Wood and Foody, 1989; Blonda et al.
1991).
The process of image interpretation holds a number of
hierarchically ordered classification steps following
from a land-cover hierarchy. An example hierarchy is
depicted in figure 3. On the first level vegetated and
non-vegetated areas are distinguished. Because the
cover types road and water can be easily derived from
topographic maps, the only non-vegetated class to be
isolated in the image is 'blond sand' Hence, the
isolation of blond sand is subject of the first
interpretation step. The second step involves the
subdivision of the vegetated area in woody and
herbaceous vegetation. The third phase deals with the
subclassification of wood into several species or species
groups. Finally, the herbaceous vegetation is
subdivided in structural types in the fourth phase.
Given the spectral and geometric resolution of the
applied image and the appearance of the land cover
types or classes in the terrain, one has to determine
whether the class has a crisp or fuzzy character.
Consider for example false colour images with a
resolution of 0.5m, than the classes wood, herbaceous
and sand are presumed to form discrete objects. Hence,
the crisp classification of the image to the three classes
deals with the estimation of a triple of membership
values (MV ^, My ***5 My?"5 for each image pixel,
where only the values 0 and 1 are allowed for no
membership and full membership respectively. For
instance the vector (0,0,1) means that the pixel is
covered with sand. The subclassification of wood to
species types is performed in a crisp classification
applying texture features.
Subsequently, we want to quantify the continuous
internal variation of the herbaceous cover class through
fuzzy techniques. The variation in herbaceous
vegetation structure is categorised in five classes hs ¢
ths, hs, -.hsj, which were qualitatively described as
prototypes (table 1). In the terrain some sites will show
a high resemblance with only one class, while others
have properties belonging to two or more of these
classes. Typically, these partial memberships can be
Quantified in a fuzzy classification, where the
resemblance of a site with a class is indicated by a
membership value MV, Ve'tl01) Hence Ya site) is
characterised by a vector of five membership values
MM Mau Mes Mn) For instance the
vector (0, 0.6, 0.4, 0, 0) indicates that the site has nearly
equal resemblance with hs, and hs, and no similarity
with hs, hs, and hs, The problem of fuzzy image
interpretation is now to estimate a vector of
membership values for each image pixel covered with
herbaceous vegetation. A detailed description of these
estimation methods captured by a the supervised
hierarchical classification system is reported by Droesen
et al. (1995).
Tab. 1. Description of the herbaceous structural
subclasses (hs) (after Assendorp and van der Meulen,
1994).
hs1: Thin grass/herb cover with blond sand
Blond sand, i.e. sand with negligible amount of organic
matter, has, by far, the largest contribution in this
coverage type. It is however accompanied by pioneer
plant types. Herbs are annual as well as biennial. Grass
types are mainly solitary and clonal which react more or
less positive to wind activity. Tussock forming grass
types can be present.
hs2: Intermediate herb/moss cover with grey sand
Largest contribution to the overall coverage is with
mosses who react more or less positive to or can sustain
some geomorphological activity. Bare grey sand, i.e.
sand with organic matter content, has a substantial
contribution to the overall coverage. Herbal plant types
are annual and biennial with locally some perennials.
Some woody plants at the sub-pixel level can occur,
grasses are solitary and tussock forming.
hs3: High moss cover
A total coverage of the soil with mosses and lichens,
very locally with annual and biennial herbs. Grasses are
nearly absent.
hs4: High moss and low grass cover
The soil is totally covered with mosses combined with a
low herbaceous vegetation. Herbs and grasses are
mainly small though larger woody plants at the sub-
resolution level can occur.
hs5: High grass/herb cover with litter
Mainly grasses and perennial or clonal herbs cover the
soil completely. The herbs are partly woody plants at
the sub pixel level. Dead ectorganic matter determines
partly the nature of this type.
217
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996