97
Figure 2: a) Boundary starting position 2; b) Latent variable p(B’),
iteration 1; c) Distance transform (DT) of class label boundaries,
iteration 1. This defines the topographic surface used for boundary
movement; d) Latent variable p(B’), iteration 2; e) Latent variable
p(B’), iteration 5; f) Distance transform of class label boundaries,
iteration 5. This defines the topographic surface used for
boundary, movement; g) Boundary position, iteration 5; h) Class
label image out; i) Class label image without use of the model.
4 Examination of model properties using a real example
Experiments were performed with data from the south west of
Western Australia shown in figure 3 below. The region includes
the area covered by the Landsat scenes of Pemberton(l 12/84) and
Mt Barker(l 11/84), and is identified as the 1:1 000 000 map sheet
si50.
Figure 3. Region covered by the MLC classifier. Area shown is
approx 400 by 200km.
The aim is to provide improved classification and boundaries for
forest plantations (Softwood and Harwood), and for native
eucalypt forests (Karri, Jarrah). Analysis of Landsat TM and areas
of known plantation and regeneration types demonstrate that •
• The ability to distinguish the different types based on
their spectra alone varies as they grow
• Mature softwood and 4-6 year old karri regrowth show
the best spectral separation from the remaining classes
• Jarrah and hardwood had the least separation
Mapped versus true class label pixel counts are given in table 1
below. These figures were used to generate table 2, P(True Class
Labels I Mapped Labels) , and table 3, P(Mapped Label I True
Class Label).
True Class Labels
Mapp
ed
SW
K
HW
J
Total
Swl
3876
178
213
292
4559
Sw 2
1216
893
379
86
2574
K1
15
5485
207
588
6295
K 2
242
19559
620
1882
22303
HW
176
1397
6973
2223
10769
J
252
1458
1572
19833
23115
B
1020
1047
4632
1165
7864
Total
6797
30017
14596
26069
77479
Table 1 - Number of pixels tabulated by mapped and true class.
P(True Class Labels 1 Mapped)
Mapped
SW
K
HW
J
Total
Swl
0.85
0.04
0.05
0.06
1.0
Sw2
0.472
0.347
0.147
0.03
1.0
K1
0.002
0.87
0.03
0.09
1.0
K2
0.01
0.877
0.028
0.084
1.0
HW
0.016
0.13
0.648
0.206
1.0
J
0.01
0.06
0.068
0.858
1.0
B
0.07
0.32
0.18
0.43
1.0
Table 2 - Probability of true class label given map label
P(Mapped 1 True Class Labels)
Mapped
SW
K
Hw
Jarrah
Swl
0.57
0.01
0.01
0.01
Sw2
0.18
0.03
0.03
0.00
K1
0.00
0.18
0.01
0.02
K2
0.04
0.65
0.04
0.07
Hw
0.03
0.05
0.48
0.09
J
0.03
0.05
0.11
0.76
B
0.15
0.03
0.32
0.05
Total
1.0
1.0
1.0
1.0
Table 3 Probability of mapped class label given true label.
The accuracy of the classification was estimated using 268 sites
having labels from the set {Jarrah, Karri, HardWood, SoftWood}.
The sites are independent from the sites used to train the classifier.