used are at a 
i.6 km by 12.6 
and are based 
i by British 
The attribute 
eparately in a 
CCRS as label 
.c information 
in Figure 4. 
.es are loaded 
1 is extracted 
ation database 
maps graphi- 
d the forest 
. The maps 
Therefore, 
based on less 
leen a limita- 
ason, LDIAS is 
ler number of 
forest classes 
epresentation. 
s used subse- 
forested and 
Ls done with 
com the MSS or 
a summary of 
-date, multi- 
i processed on 
errors and 
Mercator (UTM) 
i by acquiring 
ad calibrating 
aduced by the 
at of Energy, 
! the standard 
The map scale 
ed to a pixel 
st information 
agery. 
ally based on 
ground infor- 
» are used to 
asequent clas- 
asified within 
forest - non- 
Eiltered using 
Joldberg and 
;e is assessed 
he procedure 
:ied image is 
a grid file is 
with smooth 
DMRS database 
:es are placed 
Ltal map file, 
valuation. In 
i one of our 
as an example 
le experiment 
;e of an area 
the Kootenay 
s obtained on 
Lie on a river 
surrounded by 
sted areas are 
am with ages 
Ion-coniferous 
in any of the 
.ear-cut areas 
vary from 12 hectares to 71 hectares. The ratios of 
the boundary pixels to their total areas vary from 
18% to 43%. 
The preprocessing methods used for this experiment 
are: 
A) reflective bands - 6 TM bands used; IR band 
not used; 
B) normalized differences: 
The classification of the ratioed image was 
regrouped to bring out the separation of pine and 
mixed spruce/fir. The confusion matrix for this 
case with the maximum likelihood algorithm is given 
in Table 4. The clear-cuts, both new and old, are 
well identified, as is pine. However, the mixed 
spruce and fir class has a classification accuracy 
of only 67%. The average classification accuracy 
for the four classes was 76%. 
band i 
- 
band 
band i 
+ 
band 
* 128 + 128 
where i, j are adjacent spectral bands. 
Only four bands of normalized differences ratios 
were used as TM bands 2 and 3 are too highly 
correlated to give useful ratios. 
Two methods of training were used. The first 
training method was the traditional one of selecting 
training areas with a variable cursor on the image. 
The user judged whether the cursored area was homo 
geneous, perhaps using histogramming inside the 
training area to confirm the selection. The selec 
ted training areas were grouped into classes defined 
by the user. The classes used for the first train 
ing method were: old clear-cuts (cut areas which 
were 5 to 40 years old ); new clear-cuts (cut areas 
less than 5 years); and forest cover (uncut forest 
greater than 40 years in age). Two classifications 
were performed with preprocessing methods A and B 
given previously. 
The second method of training involved user 
selection of polygons from the GIS for all classes 
except new clear-cuts, for which we used the first 
training method since these cuts were more recent 
than the data used to make the inventory map. 
Because of the existence of polygons with complex, 
multi-modal classes, the automatic selection of 
polygons for training, without user intervention, 
failed to produce training sets which gave 
acceptable classification accuracies. 
3.3 Results 
Over 1800 polygons can exist in the digital forest 
map, but we are limited to classifying 256 classes 
or less. Hence, polygons were grouped from the 
database before classification. The polygons 
grouped into the same classes were identical in all 
of the following important attributes: 1) major 
forest species - all species which constitute 15% or 
more of the polygon; 2) age class - the weighted 
average age of the major species listed in 20-year 
stratifications; 3) site condition - the predicted 
productivity of the area at the time of surveying; 
the site condition was stratified into 4 subjective 
categories. The polygons were grouped into 63 clas 
ses, and then later grouped into the classes given 
in Section 3.2. 
Table 3 lists the average classification accura 
cies obtained for the selected classes for the two 
preprocessing methods. 
Table 3. Average Classification Accuracy 
DATA 
SET 
METHOD 
6 BANDS 
4 RATIOS 
Cursor Selection 
of Training Areas 
98.8% 
±1.2% 
97.7% 
±1.2% 
Point Selection 
of GIS Polygons 
97.4% 
±3.7% 
93.6% 
±3.8% 
This classification experiment was repeated using 
a hierarchical Logit classifier. For this 
classifier, the classification decision was broken 
down to a series of binary decisions. At each 
decision point, a binary choice probability (p) was 
computed using the following equation: 
N 
log(p/(l-p)) = a 0 + E 
i=i 
where the a^ coefficients are derived from the 
training data and X£ represents the intensities in 
feature "i". Our Logit implementation accepts up to 
256 classes and 16 channels. The confusion matrix 
for the ratioed image with the hierarchical logistic 
classifier is given in Table 5. The weighted (by 
the number of pixels in a class) average classi 
fication accuracy achieved was 87.5%. This is 
better than the maximum likelihood result, but there 
is still significant overlap between the pine and 
spruce-fir class, and the pine and new clear-cut 
class. The introduction of spatial features should 
improve this classification result. In all cases, 
the results with Thematic Mapper data were much 
better than those for MSS data. 
Table 4. Confusion matrix for the ratioed image 
classified with the maximum likelihood algorithm 
True 
Choseh\Class 
Class 
OLD 
CLEAR 
CUTS 
MIXED 
SPRUCE, 
FIR 
PINE 
NEW 
CLEAR 
CUTS 
OLD CLEAR 
CUTS 
93.0% 
8.2% 
4.1% 
0.8% 
MIXED 
SPRUCE, FIR 
4.9% 
67.2% 
14.9% 
1.2% 
PINE 
1.5% 
18.1% 
80.3% 
0.0% 
NEW CLEAR 
CUTS 
0.2% 
2.2% 
0.0 % 
98.1% 
Weighted mean classification accuracy = 75.9%±0.5% 
Table 5. Confusion matrix for the ratioed image 
classified with the hierarchical logistic 
classifier 
True 
Choseit-vClass 
Class 
OLD 
CLEAR 
CUTS 
MIXED 
SPRUCE, 
FIR 
PINE 
NEW 
CLEAR 
CUTS 
OLD CLEAR 
CUTS 
90.6% 
5.9% 
3.6% 
2.7% 
MIXED 
SPRUCE, FIR 
9.2% 
91.7% 
29.2% 
32.1% 
PINE 
0.2% 
1.7% 
67.2% 
0.0% 
NEW CLEAR 
CUTS 
0.0% 
0.6% 
0.0% 
65.3% 
Weighted mean classification accuracy = 87.5%±0.4%