In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
Kappa
OA%
OM%
CM%
disc., stand
0.74
91.1
16.1
7.2
£-NN, stand
0.72
91.5
24.3
4.5
disc., segment
0.71
90.1
13.8
8.8
&-NN, segment
0.62
86.6
18.8
11.5
Table 6. Classification results of the discriminant analysis
(disc.) and the &-NN method (£-NN) at the stand
level, stand: stands as classification units, segment:
segments as classification units. OA%: overall
accuracy, OM%: omission error, CM%: commission
error.
4. DISCUSSION
A semi-automatic method for change detection in boreal forests
using bi-temporal aerial photographs was developed. The
photographs were taken as close to each other as possible with
respect to time, date and location. The change detection was
tested at the stand and segment levels.
Considerable changes were classified, at best, without error.
The results of moderate changes were also clearly better than
achieved by some earlier studies (Saksa et al., 2003; Haapanen
& Pekkarinen, 2000). The use of correlation coefficients for
image matching at the stand level and in the classification
together with other spectral features was considered to be the
key element in obtaining the good results.
The use of the &-NN method with multi-objective optimisation
was also found to be very effective for detecting changes at the
stand level. The large number of possible features in the
distance function requires an efficient algorithm for determining
the optimal formula of the distance function. Furthermore, the
definition of the formula of the objective function is crucial for
achieving the optimal solution. For example, when the Kappa
value was the only objective variable in the optimisation, the
Kappa value of the classification of the reference data was same
as in the discriminant analysis. However, the classification
results of the Moderate-change class were then considerably
worse.
In practice, obtaining aerial photographs with the same spatial
specification is not a problem. Nevertheless, in Finland, weather
conditions limit the number of suitable days for aerial
photography. Weather conditions may therefore complicate the
acquisition of photographs with similar temporal specifications
and thus limit the operational use of the method. In this study,
photographs with similar specifications were obtained, which
enabled an exploration of how reliable change detection is when
using the bi-temporal aerial photographs in nearly optimal
conditions. Even where bi-temporal aerial photographs are
taken at identical locations, the time difference may play an
essential role; the shadows of the treetops move several metres
in a rather short time. If these shadows are captured in both
images, on the ground or some other way, the rectification
adjustment based on DNs correlations, might move the stands
so that the shadows are matched. Thus, the real objects might
not be matched in the best possible way and the classification
results would then be questionable. The effect of shadow
movement was also noticed in the study of Im and Jensen
(2005).
There are some issues that require further study. First, one
image-pair may not be enough to cover the area of interest. The
mosaicking of images introduces error, which must be
quantified in further studies. Secondly, another source of error
is the practical training data from operational databases, which
as noted in this study, contain incorrect operations. Thirdly, the
method studied should be tested with diverse image material,
for example, with images taken with no-optimal image
specifications and with images digital by origin. Fourthly, the
use of texture features in the classification, together with stand
level rectification, should be studied. Tuominen and Pekkarinen
(2005) and Hyvönen and Anttila (2006) found that the use of
textural features is advantageous in estimation and classification
procedures. The effect of different methods of classification on
the accuracy of the estimation should be studied. Heikkonen
and Varjo (2004) found that nonparametric classification
methods worked better than the parametric method tested here.
With a maximum likelihood classifier there was a strong
tendency to over classify the Moderate-change class, as was
also the case in this study.
REFERENCES
Campbell, J.B., 1987. Introduction to Remote Sensing. The
Guilford Press, New York.
Congalton, R.G. and Mead, R.A., 1983. A quantitative method
to test for consistency and correctness in photo-interpretation.
Photogrammetric Engineering and Remote Sensing, 49(1), pp.
69-74.
Congalton, R.G., Oderwald, R.G. and Mead, R.A., 1983.
Assessing Landsat classification accuracy using discrete
multivariate statistical techniques. Photogrammetric
Engineering and Remote Sensing, 49(12), pp. 1671-1678.
Haapanen, R. and Pekkarinen, A., 2000. Utilising satellite
imagery and digital detection of clear cuttings for timber supply
management. In: International Archives of Photogrammetry
and Remote Sensing, Vol. XXXIII, Part B7. pp. 481-488.
Haara, A., 2002. Kasvuennusteiden luotettavuuden
selvittaminen knn-menetelmalla ja monitavoiteoptimoinnilla.
Metsatieteen aikakauskirja, 3/2002, pp. 391-406. (In Finnish).
Heikkonen, J. and Varjo, J., 2004. Forest change detection
applying Landsat thematic mapper difference features: a
comparison of different classifiers in boreal forest conditions.
Forest Science, 50(5), pp. 579-588.
Holmgren, P. and Thuresson, T., 1998. Satellite remote sensing
for forestry planning - a review. Scandinavian Journal of Forest
Research, 13(1), pp. 90-110.
Hooke, R. Jeeves, T.A., 1961. ‘Direct search’ solution of
numerical and statistical problems. Journal of the ACM 8, pp.
212-229.
Hudak, A.T. and Wessman, C.A., 1998. Textural Analysis of
Historical Aerial Photography to Characterize Woody Plant
Encroachment in South African Savanna. Remote Sensing of
Environment, 66(3), pp. 317-330.
Hyppanen, H., 1999. Eriaikaiset ilmakuvat metsakuvioiden
muutosten tunnistamisessa. Metsatieteen aikakauskirja, 2/1999,
pp. 155-166. (In Finnish).
