can be
inuous
edium
e high
> areas
untry),
vision
ip data
e been
fusion
rand
olm.
ymated
es of
han in
of the
would
of the
t types
the T5
are the
od for
the T5
tion, a
n-wide
data is
ASES
d clear
spectral
es have
ify the
] SPOT
, 1990;
92).
| multi-
s aerial
f roads
SPOT
panchromatic data has also been extensively used for
topographic map revision in developing countries, where it has
been the only economically viable alternative for their extensive
revision needs.
Approximately 175 SPOT scenes are required to cover Sweden.
But changes do not occur everywhere at once, and high
resolution data may only be required for accurate identification
and delineation, and not for initial detection. Features with a
strong contrast to the background are detectable even when they
are many times smaller than the size of a pixel. Therefore, it is
likely that new clear-cuts and roads will initially be detectable
from lower resolution imagery.
A hierarchical system is envisioned where medium-resolution
data, for example from the new RESURS-satellite, is used for
initial detection and signalling of potential changes, followed by
identification and possibly delineation in SPOT Pan. RESURS
has a spatial coverage of 600 x 600 km and only 3-4 images
would be required to cover the country. Large coverage and
high repeat coverage (3-4 days) makes RESURS potentially
very useful for monitoring purposes. Successfully implemented,
the hierarchical approach would efficiently concentrate the use
of high-resolution data to those areas where changes are known
to have occurred.
2.2 Methods for detection and identification of revision
objects
In forested areas, recent clear-cuts and roads contrast strongly to
the forest itself and change detection can be based on spectral
characteristics alone, focusing on spectral anomalies. For
identification however, the spectral anomalies are not sufficient.
For example, roads, clear-cuts, and rock outcrops all have
similar spectral characteristics, and other information is
required to distinguish between them. In visual image
interpretation, factors such as shape and context are important
for correct identification of revision objects. Existing map
information is used to distinguish between old and new objects
of the same type. Additionally, the human vision system has an
ability to generalize and recognize patterns that are only partly
visible. It is for example possible to visually recognize a road
even if only sections of it are clearly detectable in the image.
In automated image classification, other types of data (thematic
or geometric) are often included in the classification, when
identification of land use categories on the basis of their spectral
characteristics becomes difficult. Such landuse identification is
based on knowledge about the relationships between the
landuse categories and the different ancillary data sources. The
ancillary data, e.g. digital elevation models and digital maps of
soil and geology, and knowledge are combined with spectral
information, derived form the imagery. Relationships must be
known (or hypothesised) and inferred via, for example, an
expert system engine (Middlekoop and Janssen, 1991). The
knowledge is commonly based on more or less complex rules.
These may be simple Boolean rules of the type IF <condition>
THEN <conclusion>, but there are also examples where much
more complicated rules are used. Recent expert system
development for image interpretation includes compact and
transparent rules, natural language interfaces, inclusion of fuzzy
membership functions, and symbolic approaches imitating
human aerial photo interpretation ( Leung and Leung 1993;
Srinivasan and Richards, 1993; Wang, 1994; Dymond and
Luckman, 1994).
535
Rule-based classification methods have one clear advantage
over statistical image classification techniques; it is possible to
mix data from different sources, which have different
properties. The strict requirements on Gaussian normal
distribution, which form the basis for the Maximum Likelihood
algorithm, do not apply to rule-based classification. Another
advantage with knowledge based methods compared to other
classification methods is the high transparency. Spatial factors,
such as size and shape of connected pixels (regions) or
connectivity of one region to another, can be included in rule-
based systems (e.g. Mehldau and Schowengerdt, 1990;
Johnsson, 1994).
When interpreting an image, qualitative linguistic values are
often used. Fuzzy sets can be used to mathematize these
linguistic values. Fuzzy sets have been used in remote sensing
for image interpretation and image classification (e.g. Gopal and
Woodcock, 1994; Wang, 1994).
Using classical set logic each individual pixel either belongs
entirely to a given class or does not belong to it at all. The
underlying assumption for fuzzy set theory is that the transition
from membership to non-membership is seldom a step function.
In fuzzy logic the concept of partial membership of an element
is used (Zadeh, 1965). A membership function which decides
the degree to which the pixel belongs to the given class is
calculated for each pixel. The function takes on values between
0 and 1, where 0 means fully outside and 1 means fully inside
the class.
Different types of functions can be used to express membership;
for example sigmoidal, linear and J-shaped functions. Linear
membership functions are used for all the rules in this study
(e.g. Centeno and Haertel, 1995). Figure 1 shows the linear
membership functions for three classes, namely low, medium
and high spectral reflectance. Pixelvalues in overlapping
regions between two classes will have a partial membership in
both. The membership values are determined by the functions.
Grade of
membership
u
0 >
0 low medium high 255 DN
Figure 1. Linear membership functions for low, medium and
high spectral reflectance
3. INVESTIGATION METHODOLOGY
3.1 Study area and investigation data
The study area is located in central Sweden and is almost
entirely covered by coniferous forest. Logging is intensive in
the area. The entire study area corresponds to the topographic
map sheet Ljungaverk 17G NV and covers 625 km?. A forested
test area of 33 km? , including several new clear-cuts and roads
was selected for this study. The test area includes a few lakes
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
