ul 2004
lefined,
nerated.
ing the
y:
can be
d-cover
on and
' water
domain
ure.
2 data:
sat TM
ititative
esholds
ssifying
lone is
spatial
y.
stanbul,
25172
f of the
imately
osed of
y forest
(Pinus
le tree,
e in the
for the
mainly
urfaces
ruction
At is.a
hermal
spatial
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
1/25 000 scaled forestry map sheets of Omerli Dam Lake and
neighborhood, acquired from the General Directorate Of
Forestry, provided valuable information about the forest types,
tree ages, trunk diameter and density as well as other features
like the road types, water boundaries and rivers.
The data set also includes a 1/5 000 scaled orthophoto. It proved
to be a valuable ancillary data in terms of spatial resolution (50
cm.), date of acquisition (September 1999) and the level of
detail.
3. METHOD
3.1 Contruction of hierarchy
The study area consists of both man-made and natural regions.
Forestry areas include deciduous (mostly oak groves) and
coniferous types (red-pine, black-pine). Nonforested areas, on
the other hand, are composed of orchards, shrubs, agricultural
fields and grass. Remaining regions are covered with water
(Omerli Dam Lake), roads and urban areas. So it would be a
good strategy to divide whole study area first into two classes as
vegetation and nonvegetation due to the heterogeneous nature of
it. This is the first level of the hierarchical classification.
Decomposing the area into such two classes as a first level
would help to eliminate errors arising from mixed pixels in
urban and residential regions from further levels.
The second level of proposed hierarchical classification
involves further decomposition of both vegetated and non-
vegetated areas and masking operations. Forest types,
agricultural areas, shrub and grasslands like open forest floor
would be separated from cach other using thresholds. Masking
helps to extract image pixel values which satisfy the criteria and
create new images which consist of target classes for further
analysis. Each type of region is further decomposed into smaller
regions recursively for detailed interpretation until regions
cannot be further subdivided.
The Figure 2 illustrates the structure of the proposed
hierarchical classification for the test region. It contains 3 levels
and 13 nodes.
Image
ty
Non-vegetation
poi.
Vegetation
prem
Forest Non-forest Water Non-water
Deciduous Agricultural — Grasslands Urban Transportation
Areas Areas
Coniferous
Figure 2. Proposed Hierarchical Tree Structure
3.2 Spectral Rule Generation
There are a few ways to increase the knowledge of the
researcher about the spectral properties of the features in the
TM image. Domain spectral knowledge — as described above —
is hardly sufficient to generate rules in order to progress in such
hierarchical methods, especially when seasonal variations of the
vegetation is considered, so additional information has to be
generated. This was done by examining training data containing
regions of the various types with the correct types being known.
513
The dimensionality was increased to 10, six of which
correspond to the TM bands; the remaining four were band
transforms, namely Normalized Difference Vegetation Index
(NDVI) and TM Tasseled Cap Transformations (Brightness,
Greenness, Wetness).
NDVI values, ranging from —1 to 1, stretched to unsigned 8-bit
data. Examination of training sites, in cooperation with the
orthophoto of the region, proved that a normalized difference
vegetation index threshold of 135 was a critical value for
discriminating between vegetation and non-vegetation. The
value was found by using region growing tool, and examining
the statistics generated from the DN values of the area of
interest (AOI).
Further investigation of the vegetated areas revealed that NDVI
and greenness components are good indicators of forest type
discrimination. The vegetation image which was created by
masking the non-vegetated areas from the original image, was
again decomposed into sub-levels as deciduous forest,
coniferous forest and non-forest regions using these two
indicators. Non-vegetation image was then divided into two
more sub-classes as urban and transportation.
3.3 Road Extraction
The extraction of roads from images has received considerable
attention in the past. Several schemes have been proposed to
solve this problem at resolutions that range from satellite
images to low altitude aerial images. The strategies proposed
fall into two broad categories. The work described in Gruen and
Li (1994), Heipke ct al. (1995), and McKeown et al. (1988)
deal with the semi-automatic extraction of roads. The human
operator has to select a certain number of points of the road
which is then extracted. On the other hand, the work presented
in Ruskone et al. (1994) is concerned with the automatic
extraction of roads. Most of the studies referred above used fine
resolution aerial images in information extraction about roads.
These models, despite their success with aerial imagery, can
hardly be applied to satellite images with coarser spatial
resolutions. In this study, the roads were extracted using the
vector coverages digitized from the orthophoto map. The
1:5000 scale-orthophoto of the study area proved to be an
excellent ancillary data for extraction of roads. An ARC/INFO
coverage was created by digitizing roads and used to extract the
road pixels from the image. A thematic roads image was created
using the vector coverage and these regions were extracted and
masked from the original image.
3.4 Hierarchical Classification
3.4.1 First Level: Domain spectral knowledge and spectral
classification rules obtained from training data were used as
input to the model created using Spatial Modeler Language
(SML) ERDAS IMAGINE version 8.3.1. For later applications
the software was customized and a user-friendly graphical user
interface was created using Erdas Macro Language (EML)
(Figure 3).
Figure 3. Hierarchical classification module access button
