In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
311
Figure 1. Stand (black) and segment (white) boundaries in aerial
photograph (near-infrared channel). Copyright for
the aerial photograph owned by Blom Kartta Oy
Altogether, 9 156 segments were used. The segmentation
algorithm was same as in the study of Hyvonen et al. (2005).
Altogether, 75 features both at the stand and segment levels
were extracted, of which 36 were spectral features extracted
from the original photographs and 39 were features from the
adjustment of and regression between the photographs. These
features were used in the change detection analysis.
Two different methods for change detection were tested; linear
stepwise discriminant analysis and the non-linear ^-nearest
neighbour (&-NN) method. The classification functions of the
discriminant analysis and the weights of the variables of the
similarity distance functions, as well as the number of the
neighbours of &-NN method, were estimated for the three
change classes in the training data. The multi-objective
optimisation, combined with the k-NN method, was used to
choose the weights of the variables of the similarity distance
functions, as well as the number of the neighbours of k-NN
method (Haara, 2000). The nonlinear programming algorithm
presented by Hooke and Jeeves (1961) was used to find the
combination of decision variables that minimised the objective
function. The optimised objective variable was the sum of the
Kappa value and the percentage of correctly classified
moderate-change stands. Both methods, i.e. the estimated
classification functions of the discriminant analysis and £-NN
method with the training data as the reference data, were then
applied to the test data. The accuracy of the classification was
studied at the stand level.
The accuracy of the classification results was evaluated by
means of confusion matrices (Campbell, 1987) and the overall
(OA), producer’s (PA) and user’s (UA) accuracies were
calculated (Congalton et al., 1983; Story & Congalton, 1986).
The OA can give too optimistic results if the proportion of one
class is high (Campbell, 1987) and for this reason the k (kappa)
coefficient was calculated (Congalton & Mead, 1983).
3. RESULTS
When stands were used as classification units, the overall
accuracy of the discriminant analysis was 91.1%, the Kappa-
value 0.74, the omission error 16.1% and the commission error
7.2% (Tables 2 and 6). The main interest concerned thinned
stands, of which 90.7% were found. When segments were used
as classification units and the results were studied at the stand
level, the overall accuracy of the discriminant analysis was
90.1%, the Kappa-value 0.71, the omission error 13.8% and the
commission error 8.8% (Tables 3 and 6 ). Of the thinned stands
89.7% were found.
Wfren using the k-NN method, the overall accuracy was 91.5%
and the Kappa value was 0.72 (Tables 4 and 6). The proportion
of correctly classified thinned stands was 82.5%. When the
classification was done at the segment level, the stand wise
overall accuracy was 86.6% and the Kappa value 0.62 (Tables 5
and 6). In this case, 86.6% of the thinned stands were found.
After rectification (at the image level) 12 stands had been
moved more than 5 pixels in every channel in the direction of x
and 18 stands in direction of y in the final adjustment between
the aerial photographs. Only one stand of these was in the No
change class.
Observation
No-change
Estimated
Mod.-change
Con.-change
No-change
874
68
0
Mod.-change
35
139
0
Con.-change
0
0
44
Table 2. Classification results of the discriminant analysis at the
stand level, stands as a priori classification units.
Table 5. Classification results of the &-NN method at the stand
level, segments as a priori classification units.
Observation
No-change
Estimated
Mod.-change
Con.-change
No-change
859
83
0
Mod.-change
30
144
0
Con.-change
0
2
42
Table 3. Classification results of the discriminanat analysis at
the stand level, segments as a priori classification
units.
Observation
No-change
Estimated
Mod.-change
Con.-change
No-change
900
40
2
Mod.-change
52
121
1
Con.-change
1
3
40
Table 4. Classification results of the £-NN method at the stand
level, stands as a priori classification units.
Observation
No-change
Estimated
Mod.-change
Con.-change
No-change
834
107
1
Mod.-change
41
130
3
Con.-change
0
3
41