The International Archives of the Photoyrammetry. Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beiiinp 2008
Field and Wetland Vegetation were also significant. In addition,
Paddy Field and Wetland Vegetation area, which in reality exist
only in the lowlands, and shown in the uplands as well. This is
most likely due to the fact that Paddy Field and Wetland
Vegetation patches may not have held water at the time the data
was acquired, and thus showed spectral characteristics that can
easily be misinterpreted as Bare Ground or other upland
landscape types. Using only spectral characteristics, producer’s
accuracy for Paddy Field was as low as 45%.
In contrast, classification results incorporating the GIS data
showed overall accuracy and overall Kappa index above those
for the spectral only classification. This difference was
especially evident in terms of producer’s accuracy for
Grassland, Wetland Vegetation and Paddy Field. These results
show that misclassifications among the upland and lowland
landscape types can be eliminated by using the topographic data.
Figure 8 shows an expanded view of one section of the master
landscape map (a) compared to the object-based classification
using topographic data results (b), and the object-based only
classification results (c). As can be clearly seen, the
misclassifications of lowland landscape types Paddy Field and
Wetland Vegetation as upland landscape types Bare Ground
and Grassland, have been eliminated by incorporating the
topographic data. Although the user’s accuracy for Paddy Field
was improved, that for Grassland £nd Wetland Vegetation
remained low. This is most likely due to the fact that at the
time the data was acquired some of the paddies had been
planted in a winter crop of wheat, which was easily confused
with Wetland Vegetation or Grassland.
In addition, although the producer's (78%) and user's (26%)
accuracies for Rural Residential improved slightly, the gap
between these two remained large. In both classifications, the
mean value for each object was used to derive the minimum
distance. In the Rural Residential landscape, very tiny patches
of forest, Bare Ground and Grassland are mixed together. For
this type of landscape some improvement in accuracy may be
obtained by using the standard deviation rather than the mean
value.
IKONOS image, master landscape map, and segmentation
results for a section of the target area are shown in Figures 9
and 10. In Figure 9, segmentation results for Rural Residential
are shown at two different scale parameters, SP=66 (Fig.9c) and
SP=235 (Fig.9d). As can be seen, at the lower scale parameter
each individual object is too small, resulting in over
segmentation. The higher parameter, on the other hand,
produces a much more accurate segmentation.
In Figure 10 the same comparison is made for the othei
landscape types. In this case, the higher parameter fails to
distinguish among the different types of vegetation. These
results indicate that the best scale parameter may vary
according to the type of landscape being segmented. Further
research in this direction may lead to future improvements in
accuracy.
4. CONCLUSION
This research was designed to test and improve systems for
applying object-based classification of very high resolution
IKONOS data to mapping of landscape types. In addition to the
extant multi-spectral object-based classification system as
developed by Kamagata et al (2006), topographic data derived
from field surveys and topographic maps was used to perform
an initial segmentation that divided this research area into two
topographic zones. The results showed that some of the
misclassification problems that plagued the extant system could
be eliminated by this method. Incorporation of topographic
data should prove especially useful in the Japanese countryside,
where various small patches are scattered in a complicated
mosaic pattern. The research also indicated that further gains in
accuracy can be achieved by adjusting the scale parameter to
the characteristics of each landscape type.
ACKNOWLEDEMENTS
The authors would like to thank Dr. Hijiri Shimojima of TUIS.
Professor Kevin Short of TUIS helped with the English editing
work.
REFERENCES
Baatz, M., Schape, A., 2000. Multiresolution Segmentation -
an optimization approach for high quality multi-scale image
segmentation. In: Strobl, J. and Blaschke, T. (Eds.) Angewandte
Geographische Informations Verarbeitung XII. Wichmann-
Verlag, Heidelberg, pp. 12-23.
Fujiwara, M., Hara, K., Kevin, M. S., 2005. Changes in
landscape structure of "yatsu" valleys: a typical Japanese urban
fringe landscape. Landscape and Urban Planning, 70 (2005),
pp. 261-270.
Hara, K., Kamagata, N., Ishitsuka, T., Tomita, M., 2007.
Object-based classification of rural landscapes using remotely
sensed data of various resolutions. In: Proceedings of the 7 th
IALE World Congres, Wageningen, The NETHERLANDS,
Part2, pp. 642.
Kamagata, N., Hara, K., Mori, M., Akamatsu, Y., Li, Y. and
Hoshino, Y., 2006. A new method of vegetation mapping by
object-based classification using high resolution satellite data,
In: Proceedings of I s ' International Conference on Object-
based Image Analysis (CD), pp. 5.