ture 
s an 
are 
the 
low- 
1igh- 
ng a 
right 
| the 
; are 
reby 
s for 
this 
f the 
| the 
rant. 
] out 
ower 
from 
arge 
field 
rided 
in a 
the 
the 
ique 
rcel, 
) be 
are 
on of 
ther 
lors 
2, Or 
.- In 
jy to 
s by 
nilar 
3 as 
‚the 
à in 
alid 
ount 
ss a 
from 
The 
age 
I no 
> is 
the 
> no 
ver, 
l in 
ving 
els. 
y to 
lass 
nds. 
n of 
trix) 
this 
inal 
, are 
an 
atic 
e to 
3 in 
inge 
the results of classification, particularly if the 
maximum likelihood decision rule is used as a 
classifier. Let us now look at the solutions to these 
problems provided by GEOCLASS. The first step with 
GEOCLASS is the classification of the color 
scattergram data. This will provide a much more 
accurate validation of the training performed than a 
simple estimation of probability values in 
misclassification. The classification of the two- 
dimensional scattergram representing the entire 
universe of parcels precedes the classification of the 
corresponding image. This undertaking will allow 
the analyst to verify whether the classification process 
has indeed grouped together parcels that, according to 
statistical distribution criteria (cluster analysis) and 
color, should belong to the same class. Since a unique 
code number is associated with each parcel, it is 
feasible to cross check scattergram data with imagery 
data. Experimentation has shown that valid changes 
to the number of classes and to the training samples 
associated with them can be introduced at this stage 
by this procedure. Only when the analyst is satisfied 
that the scattergram data are properly classified, the 
actual classification of image data starts. Image data 
are classified in successive passes (Steffensen and 
Smith, 1978). In the first pass each parcel (see again 
the lower right quadrant of figure 1) is classified as a 
single class. In each of the successive passes, the 
individual pixels that were filtered out during the 
segmentation process will be also classified. Before 
deciding about a pixel attribution to a class or a 
combination of two classes (boundary pixels), 
reference is made to the type of already classified 
neighbours. The logic of this process is that, although 
the signature of a pixel alone would not be necessarily 
sufficiently distinct to determine unequivocally its 
thematic nature, the proximity to an already classified 
pixel will help in making a correct decision. To 
provide an example, if a “pure” pixel has a signature 
which is similar to two different classes but all its 
neighbours belong to only one of these two classes, it is 
most probable that the pixel in question also belongs to 
the same class as its neighbours, although in pure 
probabilistic terms it may actually belong to the other 
class. However, if the sij.ature of the pixel in 
question would be such as to exclude the possibility 
that it belongs to the same class as its neighbours, the 
mere presence of neighbours having different 
attribution is not a valid reason for overruling the 
decision made. Unfortunately, post-classification 
filters widely used commercially do just this. 
9.4 Classifieation Results 
The GEOCLASS approach was tested by comparing 
the results of classifying the agricultural test site data 
of figure 1 with the results previously obtained in 1986 
by Steffensen and Mack (ref.cit.) using a conventional 
maximum likelihood pixel-by-pixel classifier. To 
overcome the difficulties inherent to the multispectral 
classification of crop types at this early stage of growth 
(crops were seeded only 30 to 60 days earlier) extensive 
use of ground truth information (a detailed field map 
recording ground cover, color, growth stage, etc., for 
more than 100 fields) was made. Actually, all training 
samples were selected within the ground truth 
sample. The comparison of the results of 
classification with the ground truth map data yielded 
an average accuracy of 87%. This level of accuracy 
would have been impossible to achieve in a 
conventional classification without the availability of 
the ground truth information. In contrast, the 
GEOCLASS approach was applied without any 
reference to the ground truth data. Ground truth data 
were only used to assess the accuracy of the results of 
classification. This assessment indicated that an 
905 
equal, if not better, accuracy was obtained for each 
crop type previously classified with the conventional 
approach, and that an average accuracy of 90% was 
achieved. 
4. CONCLUSIONS 
This paper illustrates the philosophy behind the 
development of a new approach to the thematic 
classification of multispectral imagery. Instead of 
individual pixels, agglomerates of pixels having 
similar spectral and spatial attributes are the objects 
of classification. The key factor for a successful 
implementation of this methodology is the ability to 
subdivide any multispectral image into segments 
closely related to terrain features of interest. Beside 
the successful experiment conducted with the 
GEOCLASS image analysis software on an 
agricultural test site in Manitoba, Canada, illustrated 
above, a number of other agricultural areas in 
Canada and Europe were also successfully processed, 
involving not only LANDSAT-TM but also LANDSAT- 
MSS, MOS and SPOT images. It appears, therefore, 
that this innovative approach could provide universal 
application to agriculture. GEOCLASS was also 
applied to the automated identification of individual 
tree species using large scale (one meter pixel size) 
Multispectral Electro-optical Imaging Scanner (MEIS 
II) airborne data (Mc Coll et al., 1983). The scene used 
was acquired on Oct.29, 1985 over the Petawawa 
National Forestry Institute research forest located 
near Chalk River, Ontario. Using the GEOCLASS 
segmentation algorithm, the central portions of 
individual tree crowns of Red Pine, White Pine, White 
Spruce, Red Spruce and Norway Spruce occurring in 
pure and mixed stands with open to dense crown 
closures were successfully isolated. The filtering out 
of the crown portion mostly affected by large 
variations in reflectance, i.e. the rim, eliminates an 
unavoidable source of error in pixel-by-pixel 
classification. Without having access to a precise 
ground truth that would provide species identification 
tree by tree, accuracy figures cannot be quoted. 
However, preliminary classification results indicate 
that higher classification accuracies can be obtained 
with GEOCLASS that with any other conventional 
approach. 
5. REFERENCES 
Me Coll, W.D., Neville, R.A., Till, S.M., 1983. Multi- 
detector Electro-optical Imaging Scanner MEIS II. 
Proc. 8th Can. Symp. on Remote Sensing, 3-6 May 1983, 
Montreal, pp.71-79. 
Ryerson, R., 1989. Image interpretation concerns for 
the 1990s and lessons from the past. Phot. Engn. and 
Remote Sensing, 55 (10), pp.1427-1430. 
Sijmons, K., 1987. Computer-assisted detection of 
linear features from digital remote sensing data. ITC 
Journal, No.1, pp. 23-31. 
Steffensen, R., Mack, A.M., 1986. An evaluation of 
Landsat TM and MSS data for crop identification in 
Manitoba. Proc. 10th Can. Symp. on Remote Sensing, 
5-8 May 1986, Edmonton, pp. 579-589. 
Steffensen, R., Smith, A.M., 1978. An analysis of the 
spatial and temporal distribution of surficial waters 
in the Minnedosa wetland district of Manitoba, 
Canada. Proc. 12th Int. Symp. on Remote Sensing of 
the Environment, Manila, Philippines, pp. 1015-1024.