76 
well but most of them disturb fine linear features 
in the image as roads and small creeks. 
The image will show almost no change after 4 
iterations. 
2.2 Edge extraction 
After smoothing edges are extracted from the scene by 
using the same set of patterns. If the pattern is 
defined the pixel is set to an edge (background). 
Result is an edge map with a lot of loose ends 
showing the weakness of the method. Therefore edge 
extraction through contouring is under development. 
Removing these unclosed boundaries is necessary to 
speed up edge tracking and labeling. Before cleaning 
a last attempt to save borders is performed by a 
program connecting ends not further away than 1 
pixel. The rest is lost. However, in the case of the 
SPOT data this connecting and cleaning leads only in 
a very few cases to elimination of 'significant' 
edges. This is checked with a visual interpretation 
of the image. An example of an edge extraction is 
projected over the original image in figure lb. 
2.3 Edge tracking and labeling 
The developed edge tracking program works as follows. 
Starting with a cleaned edge map processing begins in 
the upper left part of the image. Tracking starts at 
the upper left part of a segment following its inner 
border and storing the border itself. Back on its 
starting point it closes the polygon and submits it 
into a procedure that adaptively masks the region 
with a certain value. If this region is already 
masked before, in case of an inclusion, the masking 
value is set one value lower. If the region has 
inclusions masking value will be set to a higher 
value, depending on its own depth of inclusion. In 
this way pure statistics can be calculated over the 
image polygons starting on the deepest levels and 
masking the region when ready. Optionally the 
boundary will not be included in the statistics when 
set to the normal or starting masking value. In 
masking the image the size of the polygon is 
calculated. The user can define a minimum size for 
definitive storage. Figure If shows in which order 
the segments are tracked. The segments are filled 
with their sequential processing number (labels) 
modulo 256. Such a picture is very useful to find 
back the segment number after edge tracking. 
3 THE PROPERTY TABLE 
Polygons are stored in a property table with a 
signalised chaincode similar to Freeman(1961). 
Attributes in this list are: 
Region number, location of starting point in image, 
perimeter, size, rectangle coordinates, center of 
mass, rotation of the longest side of the Minimum 
Bounding Rectangle, rotation of the axis connecting 
the outmost points of a region (R0T2) , elongatedness 
(longest side divided by shortest side of MBR), 
fit(area of polygon divided by area of MBR), number 
of channels filled with statistics, inclusion index, 
ten means, medians and standard deviations and the 
number of chaincodes used. 
The values of the attributes of one region are 
given in table 1. 
The attributes have a fixed record length and are 
glued with the polygon string to a record with 
variable length. An accessory index file contains 
region number, byte offset to record and recordlength 
to provide direct access through some assembler 
routines. In this way minimum storage is required. 
Region image statistics can be applied with any 
image underlying the regions, even digitized 
documents. Region polygon form analyses make use of 
relatively simple algorithms bases on chaincode 
Table 1. Example of attributes of region number 415 
stored in a property table. For the location of the 
region see fig. lb. 
Region number: 
415 
Start points line, elem: 
160 
228 
Center (line/element 
axis): 
156.4 
249.4 
Perimeter (x30m.): 
119.7 
Size (30x30 sqm.): 
313.0 
Rotation angle (E-N-W;0-180 degr.): 
173.3 
Elongatedness : 
3.5 
Fit: 
0.5 
Inclusion index: 
100 
Normal rectangle; minli, maxli 
148 
164 
minei, maxel 
228 
273 
Number of chaincodes 
used : 
69 
Image statistics 
Chan, nr.: 1 
2 
3 
4 
Means : 42.2 
45.5 
161.1 
50.9 
Medians : 36.4 
38.0 
161.0 
49.5 
Stand. dev .: 20.5 
20.4 
8.8 
9.0 
handling(Freeman 1974). In the future another measure 
will be taken for elongatedness. The number of 
iterations to shrink a polygon will be divided by its 
size, also applied by Nagao and Matsuyama(1980). 
Irregularity can also be described by deviding the 
size of the region by that of its convex hull. 
From this table new tables or graphs can be created 
for other purposes. A region adjacency matrix is 
simply derived from the rectangle coordinates and the 
common boundaries of the polygons itself. 
Figures lc,d,e are showing subsets of the segmented 
image filled with some attributes converted to a 
greylevel. 
4 LEARNING FROM THE GROUND 'TRUTH' 
In the same way the ground truth polygons can be 
stored in a list. The ground truth consists only of 
agricultural classes listed in table 2. 
Table 2. Agricultural landuse categories, where 
NTR = number of training regions, 
NTE = number of test regions, 
NHA = number of hectares. 
category 
NTR 
NTE 
NHA 
1. Wine 
11 
99 
99.6 
2. Wheat 
11 
59 
55.4 
3. Bare soil 
6 
33 
23.6 
4. Moorlands 
8 
4 
18.2 
5. Alfalfa 
7 
10 
10.7 
6. Fallow. 
8 
19 
13.6 
Total 
51 
224 
221.1 
First the correlation is calculated of the category 
with all other attributes in the property list of the 
ground truth using a Pearson correlation. The results 
of all the 259 ground truth polygons are listed in 
table 3. 
Obviously there seems to exist no real correlation 
between the form factors and the categories. Only by 
chance there will appear inside agricultural landuse 
a relation to size, elongatedness or rotation of the 
main field axis because most of these factors are 
historically determined. Therefore including pure 
form factors in a classification will only be 
fruitfull when the classed are defined on a higher 
level as is shown by Nagao and Matsuyama(1980) i.e. 
separation of roads, rivers, buildings, woods, 
agricultural areas, etc. Undoubtedly there will be 
areas with a real correlation between elongatedness 
and class but to use them in a classification first 
Table 3. C. 
agriculture 
Perimeter, 
Elongatedm 
measuremen 
First coef: 
Lower coef: 
the correli 
PERI 
CAT .130 
.035 
CH 
CAT .228 
.000 
an expert 1 
It is ali 
'logic' ru] 
this area, 
car will b« 
lot but whi 
moorlands ( 
Leaving 1 
multispecti 
5 THE CLASS 
Previous at 
pixel class 
classifier 
percent of 
classified 
field toget 
parameters 
Each field 
multi-dimer 
from the pi 
A clustei 
neighbour"( 
with differ 
clustering 
The trair 
cluster lat 
the abilitj 
by the clus 
percentage. 
'acceptable 
the test re 
the correli 
available, 
the meaning 
classificat 
results. 
Being on 
things can 
simulated, 
areas and t 
fields. The 
table 4 are 
how good t! 
classificat 
meaning TE 
channel 1 c 
FIT (probat 
and will ir 
pixel class 
must be add 
describe 80 
rise when c 
class. Almo 
1 class. Wi 
cluster is 
result that 
performance 
(non-parame