the measure of distances of pixels from class centres, 
the sample that best represent a specific land feature 
is the one which best define the centre of such class. 
In conventional methodologies samples are either a) 
arbitrarily selected by visual inspection of a video 
display of imagery data (supervised classification) or 
b) automatically selected by implementing some form 
of clustering technique on raw data (unsupervised 
classification). Both of these techniques could possibly 
work only if color distinctions between populations of 
terrestrial objects would be sharp and no color overlap 
would exist between neighbouring classes, which is 
certainly not the case in remote sensing. In fact, in 
the majority of cases only a few radiance levels 
separate a class from the next. Even a skillful image 
analyst would find very difficult, if not impossible, to 
outline visually on a video screen samples that will 
define the actual location of class centres, so that valid 
statistical discriminating functions can be applied. 
Moreover, if variance values are also computed from 
these samples, as normally done with the maximum 
likelihood decision rule, the wrong assumption is also 
made that the sample variance, which in the majority 
of cases depends mostly on local factors (e.g. the 
presence of discontinuities in the canopy of crops), 
does somehow reflect the actual variance of an entire 
class of objects. As to the application of clustering 
techniques, the presence of *mixture" pixels together 
with the color overlap existing among various classes 
makes it very difficult to find valid histogram peaks, 
i.e. peaks that do represent class centres. 
3. GEOCLASS APPROACH 
3.1 Image Segmentation 
Let us look now at the solutions to these problems 
provided by the GEOCLASS approach. First of all, it 
was found necessary to isolate object populations from 
each other through a segmentation process based on 
the extraction of boundaries that delimit regions of 
homogeneous color. This process is done in stages 
and is illustrated by Figure 1. The upper left quadrant 
depicts a 4.5x3.5 kms agricultural test site located in 
Manitoba, Canada. This site .vas used for an in-depth 
evaluation of the capabilities of the TM sensor to 
discriminate a variety of field crops: potatoes, flax, 
cereals, peas, rapeseed (canola), etc. (Steffensen and 
Mack, 1986). This enhanced portion of a TM scene 
gathered on July 1,1984 was produced using band 4: 
red, band 7: green, band 2: blue. Fallow and potatoes 
appear as green tones, cereals as red tones, peas and 
canola as pink tones, woodlands as dark red tones. 
The upper right quadrant shows the results of 
extracting gradient values for each pixel. This is done 
by measuring a non-directional gradient value 
occurring within a 3x3 cell centered on each pixel. 
This process is applied band by band. If a pixel is 
surrounded by pixels of the same value in a specific 
band it will have a zero gradient value in that 
particular band. The higher the difference between 
the center pixel and its eight neighbours, the stronger 
the gradient value. The end result of this automated 
computation is a subimage in which dark areas are 
areas of color homogeneity and bright lines mark the 
color changes occurring within the scene. It can be 
seen on figure 1 that high gradient values affect a few 
pixels at the edge of each agricultural field and that 
different gradient values occur in different bands, 
depending on the crop types involved. Yellow lines 
(high gradient values in bands 4 and 7) mark the 
boundaries between exposed soil and vegetation, while 
purple lines (high gradient values in bands 4 and 2) 
mark boundaries between different green crops. To 
further understand how the upper right quadrant of 
figure 1 is related to the upper left quadrant, one can 
904 
focus his attention on the triangular feature 
appearing on the upper left quadrant. This is an 
abandoned airport. Since the runways are 
represented by more than one pixel in width the 
gradient algorithm is capable of isolating a black low- 
gradient pixel having on both sides bright high- 
gradient pixels (see right side). Now, by applying a 
ridge-edge extraction algorithm to the upper right 
quadrant we obtain the subimage illustrated in the 
lower left quadrant, where precise boundary lines are 
defined. This process is an iterative process whereby 
the image analyst chooses the proper thresholds for 
achieving satisfying results. The final step in this 
segmentation process is an automated filtering of the 
boundary lines of the lower left quadrant from the 
upper left quadrant to obtain the lower right quadrant. 
A zero value in all bands is assigned to the filtered out 
pixels resulting in the black areas shown on the lower 
right quadrant. Notice that the parcels resulting from 
the segmentation process do follow in the large 
majority of cases in shape and size the different field 
crops and that only rarely individual fields are divided 
in multiple sub-units. 
3.2 Training 
Assuming that each agglomerate of pixels within a 
parcel portrays a single object, which is normally the 
case, the segmentation process provides us with the 
object population needed for classification. A unique 
identification number is assigned to each parcel, 
allowing for the average color of parcels to be 
computed in each band. Then these vector data are 
displayed as a color scattergram on the video screen of 
an image analysis system. The analyst can either 
assign to the points in the scattergram identical colors 
to those of the corresponding parcels in the image, or 
any transformed color (e.g. ratios between bands). In 
this way the analyst is provided with the capability to 
a) identify visually the location of class centres by 
grouping together scattergram points having similar 
color; and b) define valid training samples, by 
selecting parcels located around class centres as 
being representative of each class. In summary, the 
basic advantage of the GEOCLASS approach in 
training is that it allows the identification of valid 
locations for class centres, which is the paramount 
factor for a successful classification. 
3.3 Classification 
In the last phase of the classification process a 
classifier is applied to extend the classification from 
the training samples to the entire image. The 
conventional approach is to carry out this final stage 
as if the structural context of each pixel would be ot no 
significance. In other words, if an image is 
scrambled, or if we would change arbitrarily the 
relative position of the pixels, there would be no 
impact on the classification results. However, 
contextual considerations can be quite helpful in 
finding the correct classification for pixels not having 
a distinct signature and for boundary pixels. 
Boundary pixels should not be processed similarly to 
“pure” pixels. In a conventional system class 
validation is done purely on theoretical grounds. 
Valid classes are those for which the computation of 
certain statistical parameters (e.g. confusion matrix) 
indicates no class overlap in feature space. In this 
case, a high level of accuracy is expected in the final 
classification results. However, even samples that are 
not overlapping in feature space may lead to an 
unsatisfactory classification in terms of thematic 
accuracy, if these samples are not located close to 
class centres. Even relatively small changes in 
selecting training samples can significantly change 
EN CD | ch E sou Ee obe4 pie Ey (ho B PS PS LI a py es AN Mm mde pu m uk Um. Iun Gc o d 1. qu. 707577 
rt pad Co uri Pal 
CN ONTO