The International Archives of the Photogrammetry. Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beiiing 2008
Figure 1. Location of study area
2.2 Methods
To verify accuracy of the results, a master landscape map was
created based on field surveys and aerial photographs (Figure 4).
Taking into consideration the local topography and land-use
patterns, the following 11 landscape types were identified.
1. Conifer Plantation
2. Evergreen Broad-leaved Forest
3. Deciduous Broad-leaved Forest
4. Bamboo Grove
5. Grassland
6. Wetland Vegetation
7. Paddy Field
8. Bare Ground
9. Rural Residential
10. Urban Residential
11. Open Water
IKONOS data (Japan Space Imaging - multi-spectral
resoulution 4 meter, panchromatic resolution 1 meter) acquired
on 23 April 2001 was utilized (Figure 2). In addition,
anticipating the role that topographic elements would play in
the analyses, the local topography was surveyed in the field,
and a topographic map was traced to form a GIS data base
showing the boundary between the valley bottom lowlands and
the slope (Figure 3). This division between the lowlands and
the uplands (including the valley slopes), was incorporated in
the initial process of segmentation and object-based
classification.
Figure 2. IKONOS true color image of study area
Figure 3. Lowland boundary based on topographic data
Red line: the boundary between valley bottom and slope
HUla Conifer Plantation
Mi Evergreen Broad-leaved Forest
Deciduous Broad-leaved Forest
Bamboo Grove
I I Grassland
|U&li Wetland Vegetation
I 1 Paddy Field
] Bare Ground
I i Rural Residential
lim Urban Residential
HH Open Water
Figure 4. Landscape map (Master map)
Blue area shown in expanded view in Fig 8
This classification employed the system developed by
Kamagata et al (2006), using Definiens Ver.5 software
(Definiens). Initial segmentation was a muli-resolution,
bottom-up system based on the method of Baatz and Schape
(2000). The panchromatic data was used only in the
segmentation processing. In object-based classification, object
size, shape and other parameters can be adjusted to fit the needs
of the research. Texture and color of the image data were used
to classify each unit, and integration of areas was accomplished
by increasing the scale parameters. A scale parameter of SP=66
was decided on. The study area was divided by segmentation
processing, and each segment identified was considered to be
one object.
This research also integrated topographic data into the
classification. The results of the above segmentation process
were first classified into lowland and other terrain types based
on the topographic data, using a higher scale parameter. Based
on these classification results, landscape types were divided
into three categories; those found only in the lowlands; those
found only on the uplands, and those found on both.
Classification criteria were established using the decision tree
shown in Figure 5. Each landscape type was allocated into the
hierarchy correlated with the proper terrain element. Landscape
maps, field surveys and aerial photographs were used to
establish the ground truth and set training data; and the
classifications were implemented using the nearest neighbour
method based on the mean value of each object.