1045 
ar interpolation 
rtant to achieve 
really fit into 
by cubic convo- 
that is superior 
tipling a multi- 
artant to do the 
is considered to 
is important to 
Jm of changes of 
The following 
nultidate image 
ie geometrical 
ground control 
-f necessary, in 
Lynomial trans- 
> a map projec- 
have poor geo- 
ynomial trans 
image is then 
.th a minimum of 
the image, it 
control points 
sample outside 
der polynomial 
impression of 
ch point may be 
/ a high order 
ied. 
.g. classified 
resampled by 
hm involves an 
e the nearest 
ata recorded at 
like the SPOT, 
rocedures for 
lal resampling, 
netry descrip- 
and opens new 
:ion based on 
Lstics (Moccia 
>dern remotely 
leed for data 
>mmonly applied 
r sis (PCA) , or 
iim of PCA in 
' in the multi- 
many studies 
s are highly 
to this high 
ensors 
contain the 
riance in the 
component will 
the remaining 
of variance 
in the first 
31sson 1985). 
ament in e.g. 
ate data sets, 
by 50 % per 
a risk that 
ransformation. 
3 for how the 
information content will be treated are impossible to 
work out. 
An alternative method is the canonical transfor 
mation, which maximizes the separability of defined 
classes and minimize the variance within classes 
(Schowengerdt 1983). Maxwell (1976) showed very 
promising classification results using canonical 
transformations of Landsat MSS data for rangeland 
vegetation mapping. 
5. SATELLITE REMOTE SENSING, IMAGE ANALYSIS 
The digital image from e.g. Landsat consists of an 
enormous amount of information. This information can 
be divided into three different kinds of image 
elements (Haralick & Shanmugam 1974): 
1) spectral information; relates to the band-to-band 
variations of a single pixel in a multispectral 
image, 
2) textural information; relates to the spatial 
distribution of pixel values. Texture is often 
measured in a local neighbourhood of a pixel, 
3) context information; relates to an a priori know 
ledge of the surrounding of a pixel or region. 
The latter two kinds of information are often 
denoted as spatial information. A major difference 
between manual interpretation of imagery and 
automatic analyses of digital data is the use of 
spatial information. Conventional computer classifi 
cation is entirely based on spectral information in 
the location to be classified. Manual interpretation 
is, on the other hand, to a large extent dependent on 
textural, and even more, on context information. This 
means that the interpreter does not only consider the 
point to be classified, but use also the surrounding 
for interpretation. The human way of taking advantage 
of context information will always be superior to 
what machines can do. 
Spectral information, on the other hand, is very 
efficiently handled by computers, and often superior 
to manual interpretation, both in terms of accuracy 
and speed. A semi-automatic approach to remote 
sensing can therefore be the optimal way to analyse 
complex data. In applications where quantitative 
results are desired, it is, however, necessary to 
work with computerized analysis of digital data. 
5.1 Multispectral classification 
Multispectral classification has been one of the most 
common remote sensing methods. Among the classifica 
tion algorithms, the maximum likelihood (ML) one has 
been most popular. All classification algorithms have 
in common that each class is assumed to be spectrally 
unique. Many authors have criticized this assumption. 
Graetz et.al. (1982) wrote "... the underlying 
assumption of both these approaches (supervised and 
unsupervised classification), i.e. informational 
classes were readily separable in spectral data 
space, was most unlikely to hold in the rangelands". 
Coiner (1980) also rejected classification as a 
reliable method, "Conventional methods (i.e. multi 
spectral classifications) were felt to incorporate 
unnecessary local variability (noise) that submerged 
information relevant to development of regional 
vegetation change and status mapping". 
Although multispectral classification has several 
weaknesses and is much criticized, it is in many 
cases necessary and desirable to use the technique. 
Several ways to improve the methods exist. 
1) Prior probabilities in ML-classification. The 
expected areal distribution of classes can be used 
as prior probabilities to improve the result of 
ML-classifications. Commonly the prior probabili 
ties are assumed to be equal. If the expected 
areal frequencies of classes can be estimated from 
external data, it can be used as separate prior 
probabilities, to improve the ML-classification 
(Schowengerdt 1983, Bauer et.al. 1979 and Strahler 
1980) . 
2) Use of ancillary data. Non-spectral bands, consis 
ting of e.g. elevation data (DEM), soil map and 
climatic data, can be incorporated in the classi 
fication procedure in several ways. The calculated 
probabilities in the penultimate step of ML- 
classification (before final class labelling) can 
be modified by a non-spectral band (Strahler 
1980). A post classification relaxation technique 
was applied by Richards et.al. (1982), to incorpo 
rate a DEM in a forest classification. 
The use of non-spectral bands as extra features in 
conventional ML-classification should, however, 
mostly be avoided. There are several drawbacks 
(Richards et. al. 1982): 
- the ancillary data must be normally distributed, 
which means that categorical types of data can not 
be used, 
- the relative scaling of the spectral and the ancil 
lary data may be important, 
- more training data may be necessary to give 
reliable covariance estimates, 
- there is a quadratic increase in classification 
cost with the addition of features when using Gaus 
sian ML classification 
A simple but efficient way of using categorical 
data as non-spectral bands is to carry out conven 
tional ML-classification under masks (stratifi 
cation) . Areas can either be excluded by the use of 
thematic bands, or different sets of training 
statistics can be used within different regions. 
However, a constraint to stratification is that the 
technique is deterministic, i.e. there are no 
gradations or fuzzy boundaries between mapped classes 
(Hutchinson 1982). 
Spatial information can be used in classification, 
either as nonspectral bands or that some spatial 
operation be included in the classification 
algorithm. The standard deviation from a 3x3 filter 
as a non-spectral band was used by Strahler et.al. 
(1978 & 1979). This was reported to significantly 
improve the classification accuracy of tree species 
compared to conventional ML-classification. 
Alternative classification algorithms include 
contextual classifications. Here the classification 
algorithm considers not only the spectral properties 
in a single pixel, but also the spatial context in 
which this pixel is located. An important development 
of contextual classifiers is the so called per field 
classification in agricultural applications of remote 
sensing (Bauer et.al.1979). Instead of classifying 
each pixel separately a larger homogeneous region, (a 
field) is classified by using the mean signature of 
this region. One way of doing this has been to use 
the geographical coordinates of each field. It is 
however unpractical to be dependent on this kind of 
information, which has to be updated regularly. To 
overcome this problem an image partitioning algorithm 
can be combined with a sample classifier for the 
automatic delineation of homogenous areas (Kettig & 
Landgrebe 1976). 
A test of different classification algorithms, 
contextual as well as non-contextual ones, was 
carried out by Lid Hjort & Mohn (1984). The 
contextual algorithms were found to be superior to 
the non-contextual one, concerning classification 
accuray. Though the computer time needed were between 
5 and 25 times higher for the contextual algorithms. 
A constraint to multispectral classification is the 
very time consuming procedure, especially when using 
the new high-resolution data from SPOT and Landsat TM 
for multi-temporal applications. In the case of 
ML-classification the computer time increases