527
I— *
I CONTROL
-I NOMENCLATURE TEST
CLASSIFIEO IMAGE
ATISFAÇTORV NOT.SATIS
detection of the
igital classifi-
a supervised met
ier .
els around vege-
Boundary pixels
represented port-
ypes. Their values
area of each
the relative ref-
. This was also
es, and this was
gy, heterogeneous
over.
ion were checked
with the digital
etation was asses-
as within the
al-Cavarre and
e 2) .
[ood mapped by the
SPOT-HRV) with a
•T-simulation
it 5 TM (treshold:
idsat 5 TM-sensor
vegetation cover
ild: 3.001. H2 Dry
for the SPOT-simu-
reshold: 5.00),
bad detected and
itercontent and the
:ypes of vegetation.
?ed a very good
ition (treshold:
3.00) and 97 % for the Landsat 5 TM (treshold: 5.00) ,
respectively.
3.4.3 Forests
G7' Rich meadow with willow and birch showed a very
good accuracy for the SPOT-simulation (accuracy of
91.3 %). What G7 Birch forests (meadow type) concern,
the accuracy was rather low. The same low percent of
accuracy were shown by E5b Grey alder forests (poorer
type), Bl Birch forests (richer heath type) and A4b
Birch forests (shrub type) were rather bad mapped in
the area of Dividal-Cavarre by the both systems (ac
curacy up to 24 %). In the area of Saratr0a-Habafjell
was this vegetation cover type mapped by a accuracy
for SPOT-simulation of 87 % (treshold: 3.00) and 93 %
for Landsat 5 TM (treshold: 3.00), respectively.
E5a and E5c, Grey alder forests, were very good
mapped by the both systems, with a accuracy of 94 %
for SPOT (treshold: 3.00) and 88 % for Landsat 5 TM
(treshold: 3.00), respectively.
Table 1. ACCURACY OF THE DIGITAL CLASSIFICATION
Dividal - Cavarre Accuracy in percent
Covertype
Treshold
SPOT-sim
3.0 4.0 5.0
Landsat 5 TM
3.0 4.0 5.0
Vegetation-
map
(ha)
A4b Birch
forests
7.6 7.8
7.8
7.0 7.6 7.8
31.9
(heath)
23% 24%
24%
21% 23% 23%
Bl Birch
forests
2.1 1.8
1.6
1.6 1.5 1.5
6.4
(richer heath)
32% 28%
26%
25% 23% 23%
E5a,c Grey Alder
forests' (very rich
4.8 5.5
5.6
5.2 5.9 5.9
4.6
type)
94% 84%
81 %
88% 78% 78%
E5b Grey Alder
forests (poorer
5.5 5.8
5.6
8.0 8.1 8.1
1 .6
type)
29% 27%
26%
20% 19% 19%
G7 Birch
forests (meadow
1.2 1.2
1 .2
11.2 11.6 11.9
0.2
type)
16% 16%
16%
1% 0.7% 0.7%
H1 Extremely
24.0 28.
0 29.8
9.2 10.1 10.4
15.8
dry shrub
66% 56% 53%
58% 63% 66%
H2 Dry
14.4 16.
2 17.6
47.9 51.3 51.9
27.6
Shrub
52% 59% 64%
57% 53% 53%
H7 Rich
6.2 7.2
7.2
1.6 1.4 1.4
1.8
shrub
29% 25%
25%
89% 78% 78%
A1/A2 Farmland
0.6 1.0
32% 50%
1 .7
55%
0.4 0.3 0.3
20% 15% 15%
2.0
Unclassified pixels
Snowcover
23.6 13.
9.6 11.
6 10.6
5 11.0
7.4 1.7 0.3
3.4.4 Farmland
AA1/AA2 Farmland was very bad detected and mapped by
both of the sensorsystems of SPOT-HRV and Landsat 5 TM,
due to the early phenological stage and the high water-
content in the soil.
3.4.5 General comments and discussion
The number of unclassified pixels were rather great,
due to the early phenological stage with wide varia
tion within the vegetation cover types and the distri
bution of snowcover.
The investigation has shown that the Landsat 5 TM-
scene from the springtime almost had the same accuracy
by supervised classification as the SPOT-simulated
imagery, due to the better radiometric resolution for
the TM-sensor compared to the simulated HRV-sensor on
the SPOT-satellite, for mapping purposes.
By comparing the classifcation results with the
"ground truth" - digital map, the main trend was that
Table 2. ACCURACY OF THE DIGITAL CLASSIFICATION
Saratr0a - Habafjell Accuracy in percent
Covertype
Treshold
SPOT
3.0
-sim.
4.0 5.
.0
Landsat 5 TM
3.0 4.0 5.0
Vegetation-
map
(ha)
Va Water
2.4
2.5 2.
.6
2.5
2.6 2.7
2.7
(lakes and
89%
92% 96%
92%
96% 100%
rivers)
A4b Birch
forests
27.1
35.0
37.0
25.6
33.0 34.1
23.8
(heath)
87%
68%
64%
93%
72% 69%
G7 1 Rich meadow
with willow and
2.1
2.1
2.1
2.3
birch
91 %
91 %
91 %
H1 Extremely dry
3.5
3.3
3.4
3.0
3.8 4.1
4.4
shrub
89%
76%
78%
68%
86% 93%
H2 Dry
8.7
10.0
11.1
11.1
15.3 20.3
30.7
shrub
28%
34%
36%
36%
49% 66%
P2 Wet
5.5
6.8
7.5
0.4
0.1 0.1
3.3
shrub
62%
48%
44%
12%
3% 3%
Q4 Poor mire
15.2
16.1
17.3
12.0
11.6 12.4
12.7
(intermediate type)
83%
77%
73%
94%
91% 97%
Q5 Rich
3.2
3.7
4.3
0.7
1.3 1.4
2.0
mire
62%
54%
46%
35%
65% 70%
Unclassified pixels
33.4
16.0
10.8
42.8
30.6 22.5
the vegetation cover types which were most phenologic
al develloped, showed the best accuracy in the super
vised classification. This is due to the very early
phenological stage and the distribution of snowcover
in the mountain areas. But even the classification of
the vegetation cover types HI Extremely dry shrub and
Q4 Poor mire (intermediate type) in the mountain area
was successful, with an overall classification accura
cy of 90 % or more.
Classification of the SPOT-simulated imagery showed
that the vegetation cover types within smallareas, were
better detected and mapped due to the better spatial
resolution compared to the TM-sensor on the Landsat 5
TM satellite.
The autumn-scene taken by the Landsat 5 TM-sensor
could not be used as a basis for supervised digital
classification, due to the the very low sunelevation
and to the very late phenological stage.
Several authors have discussed the possibilities of
using SPOT- and Landsat-imagery as a basis for mapping
of vegetation. Sadowski & Sarno (1976) stated the con
ventional ML-classifier was too bad for mapping pur
poses, and they used a contextual ML-classifier ins
tead. The result of this was an improvement of the ac
curacy for the classification, and this was also the
trend for channels with a spatial resolution > 32 m.
Teillfet et al. (1981) compared MSS-data with simula
ted TM-data, and stated that the accuracy of classi
fication was improved by using TM-data instead of MSS-
data (MSS: 67 % of accuracy and TM: 83 % of accuracy)
in classification of forest cover types. This improve
ment was a result of the better radiometric and spatial
resolution for the TM-sensor. In addition they found
none significant improvement as a result of improve
ment in spatial resolution alone. This is also my ex
perience .
Jaakkola (1985) stated that it is obvious that SPOT-
data will benefit from the use of multi-point (contex
tual methods) instead of single-point classifiers. He
also stated that texture should be used in the classi
fication. I will agree in this statement what SPOT-data
concern, but for TM-data with the improved radiometric
resolution, it is my meaning that multi-point classi
fiers only can give marginal improvemant. Jaakkola (19
85) also stated that forest cover type-classification
into six classes was successful with an overall classi
fication accuracy of 90 percent or more. For some of
the forest cover types and even other vegetation cover
types, I got a successful result with an overall clas
sification accuracy of 90 percent or more. Especially