egion number 415 
location of the 
415 
228 
,4 249.4 
119.7 
313.0 
: 173.3 
3.5 
0.5 
100 
i 164 
3 273 
69 
4 
.1 50.9 
.0 49.5 
.8 9.0 
ire another measure 
le number of 
be divided by its 
:suyama(1980). 
by deviding the 
:onvex hull, 
iphs can be created 
mcy matrix is 
:oordinates and the 
tself. 
:s of the segmented 
converted to a 
ilygons can be 
consists only of 
le 2. 
Tories, where 
d of the category 
roperty list of the 
lation. The results 
ns are listed in 
o real correlation 
ategories. Only by 
ricultural landuse 
or rotation of the 
ese factors are 
including pure 
ill only be 
ined on a higher 
suyama(1980) i.e. 
ings, woods, 
ly there will be 
reen elongatedness 
issification first 
Table 3. Correlation tests on fields with their 
agricultural class. Explanation: CATEGORY versus 
Perimeter, Size, Rotation angle as described above, 
Elongatedness, Fit in MBR, a second rotation 
measurement, and the means of Channels 1 to 4. 
First coefficient = Pearson coefficient, 
Lower coefficient = the significance probability of 
the correlation. 
PERIM 
SIZE 
ROTI 
ELONG 
FIT 
CAT 
.1308 
.0353 
.1179 
.0580 
.0252 
.6861 
.0699 
.2625 
.0142 
.8197 
CHI 
CH2 
CH3 
CH4 
ROT 2 
CAT 
.2286 
.0002 
.3271 
.0001 
.3164 
.0001 
.3268 
.0001 
.0760 
.2229 
an expert has to be consulted. 
It is also almost impossible to create a spatial 
'logic' rule of neigbourness with these categories in 
this area. A bridge will be connected to a river, a 
car will be adjacent to a road, house or a parking 
lot but what will be the neigbours of a wine field, 
moorlands or alfalfa? 
Leaving the spatial 'reasoning' only the 
multispectral properties of the segments are left. 
5 THE CLASSIFICATION 
Previous attempts by Megier(1984) with the pixel by 
pixel classification using a maximum likelihood 
classifier showed that a non weighted average of 51.3 
percent of the pixels in the test areas could be 
classified correctly. Having the 'contours' of each 
field together with all the image statistical 
parameters a per-field classification can be applied. 
Each field is represented by a vector in a 
multi-dimensional space created by all the attributes 
from the property list. 
A clustering algorithm using the "mutual nearest 
neighbour"(1974) is applied to 1232 regions. Tests 
with different variables have been carried out 
clustering down to 40 nodes. 
The training polygons are used to assign the 
cluster labels to one of the six classes. Furthermore 
the ability to separate the training regions itself 
by the clustering procedure is tested and given in a 
percentage. Only when the training regions have an 
'acceptable' level of identification the results of 
the test regions have a meaning. In contradiction to 
the correllation test of form-parameters, having them 
available, some clusterings have been made to confirm 
the meaningless of these attributes for agricultural 
classification in this area. Table 4 gives the 
results. 
Being on the slippery path of statistics a few 
things can be noticed. An ideal situation is 
simulated, that is starting with only a few training 
areas and testing the performance on many known 
fields. Therefore the most important results in 
table 4 are the columns TR and TE. Column TR shows 
how good the training fields are chosen for the 
classification in this scheme. The higher TR the more 
meaning TE has. The results show that the median of 
channel 1 or 2 with channel 3, eventually added with 
FIT (probably accidental), give the best performance 
and will increase the accuracy compared to the per 
pixel classification with 15 percent. To this result 
must be added that the median of channel 2 and 3 can 
describe 80.1 percent of the test regions. Problems 
rise when clusters must be assigned to a ground truth 
class. Almost every cluster will appear in more than 
1 class. With a simple Bayesian decision rule a 
cluster is relabeled with a class. This leads to the 
result that the biggest class will get the best 
performance. This is not satisfying. Other 
(non-parametric and parametric) classifiers must be 
Table 4. Results of clustering segmented image, where 
ME = mean of percentages in all classes of correct 
classification 
OV = correct classification in perc. of train and 
test regions 
TR = perc. of correct classification of training 
regions 
TE = perc. of correct classification of test regions 
LO = num. of classes not possible to assign to a 
ground truth class 
CL = num. of clusters not possible to assign to a 
ground truth class 
IN = variables used to create vectors 
- = no sense in calculating because ground truth 
classes are lost after clustering 
ME 
OV 
TR 
TE 
LO 
CL 
IN 
- 
- 
62.8 
28.0 
1 
5 
ROTATION 
- 
- 
55.7 
37.1 
1 
13 
ELONGATEDNESS 
20.6 
45.2 
57.7 
42.1 
0 
7 
FIT 
20.1 
40.0 
60.4 
32.8 
0 
14 
ROT,ELONG,FIT 
40.7 
66.2 
76.2 
63.7 
0 
10 
FIT, MED. CH 2,3 
32.7 
60.0 
67.5 
57.0 
0 
16 
ROT,ELONG,FIT 
43.4 
66.0 
77.0 
62.7 
0 
10 
MEDIANS CHAN. 2,3 
42.1 
67.4 
76.3 
64.0 
0 
9 
MEDIANS CHAN. 1,3 
40.3 
63.4 
78.1 
58.6 
0 
9 
MEANS CHAN. 2,3 
applied to the data. Together with a more precise 
choice of the ground truth classes and a better 
segmentation procedure the results will probably 
improve. 
6 SOMETHING ABOUT EXECUTION TIME 
All programs are developed on a VAX 785 mini-computer 
in VAX Pascal with some calls to Fortran subroutines 
(mainly image file I/O). System management allows 
tasks with a size of 4 Mbt. leaving 4 Mbt. for the 
system and overhead. The image consists of 232 by 216 
pixels. All mentioned times are execution times based 
on a single use of the system. 
The edge preserving smoothing uses 40 seconds per 
iteration. 
Edge extraction, connecting and cleaning uses 5 
minutes. 
The edge tracking and region storage program works 
with a 3 by 3 subimage following the inside of a 
region. This requires a buffer of 3 adjacent lines. 
Following the edge this buffer has to be updated by 
changing the index numbering of the buffer or reading 
in a new line on a not used part of the buffer 
depending on its direction. The number of I/O's the 
program has to perform depends on the number of 
intersections of edges with each scanline. This is 
very time consuming. Therefore two versions of this 
program are written. The first program does actually 
the I/O from disk n-times. The second program reads 
the edge map into memory and copies the line number 
into the wanted place of the buffer (memory to 
memory). The advantage of the first method is that 
the program can be run virtually on every computer. 
The second version needs much more memory space but 
reduces edge tracking time from around 50 minutes to 
10 minutes execution time. The number of tracked 
fields is about 3000. 
REFERENCES 
Devijver, P.A. 1974. On a new class of bound on 
Bayes risk in multihypothesis pattern 
recognition IEEE Trans, on Computers, vol.70. 
Freeman, H. 1961. On the encoding of arbitrary 
geometric configurations, IRE Trans, vol. EC-10, 
p.266-268, June 1961. 
Freeman, H. 1974. Computer processing of line 
drawing images. ACM Comput. Survey, vol.6, 
p.57-79.