International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
There were some different methods in gathering the reference
spectra and these methods were tested and their results were
compared. First, the reference spectrum was chosen from the
reflectance of individual and exactly specified image pixels.
Next, the reference spectrum was calculated from pixels of the
training sites. Average, median and mode methods were used to
calculate the reference spectrum.
2. METHODS
2.1 Classification algorithms
2.1.1 Spectral Angle Mapper (SAM): Reflectance
spectrum of individual pixel may be discussed as an n-
dimensional vector, where n is the number of image channels.
Each vector has certain length and direction. The length of the
vector represents brightness of the target while the direction
represent spectral feature of the target. Variation in illumination
mainly effects changes in the length of the vector. Therefore,
classification is based on the direction of the vector. (Kruse et
al., 1993)
$
pe
Le
bd S,
g €
i %
Band i 2
gr
s ©
e
2
"dark point" Band j
Figure 1. The spectral angle between material A and B in two
channels case (Kruce et al., 1993).
Classification is done by comparing the spectral angles (Figure
1) between the reflectance spectrum of the classified pixel and
the reference spectra obtained from training data or spectral
libraries. Each pixel will be assigned to the class according to
the lowest spectral angle value.
2.1.2 Spectral Correlation Mapper (SCM): SAM cannot
distinguish between negative and positive correlations because
only the absolute value is considered. SCM is generated as an
improvement on the SAM. SCM algorithm is very similar to
SAM. The difference between the algorithms is that SCM
standardizes the vectors of the reflectance spectrum before it
calculates the spectral angles. (Carvalho and Meneses, 2000)
2.1.3 Spectral Unmixing: An area assigned by a single
pixel of remote sensing image usually contains a lot of different
materials. These materials are mixed together and the pixel
reflectance observed by sensors is a combination of reflectance
of individual materials. To get more information from a single
pixel the proportions of these materials can be approximated
using a spectral mixing model (Boardman, 1994). Using
Spectral Unmixing model the mixed pixel can be reconstructed
from known spectra in the image or the mixed pixel can be
divided into components.
84
2.2 Test site and field measurements
Aerial measurements with an AISA airborne imaging
spectrometer were made at Lammi in southern Finland. The
area contains mainly lakes, rural areas, cultivated fields and
coniferous and deciduous forests. Size of the test area was about
50 kilometres long and 2 kilometres wide.
Geological Survey of Finland had done fieldworks in 1999 and
2000. Training area inventory included approximately 250
training areas. For example, the primary sources of reflection,
the vegetation and the soil class were specified in field
inventory.
2.3 AISA hyperspectral data
AISA data was acquired from an aeroplane on September 1999.
Six strips were flown and raw data was gathered to the actual
hyperspectrum image and geometric correction was performed.
Finnish Forest Research Institute preprocessed the data.
Figure 2. A part of the AISA image (Ruohomiki at al., 2002).
The pixel size of the image was 1.1 meter. Image contained 17
visible and near infrared channels. The weather took a turn for
the worse during the flight and the illumination was low in last
strips (Figure 2).
2.4 Gathering the reference spectra
Classification algorithms like SAM use reflectance spectra as
reference data of classes. Reference spectra are measured from
pure and single image pixels or larger training areas. The
quality of reference spectra is an important factor that defines
the classification results. Therefore, different methods were
used in gathering the reference spectra. Besides reflectance
spectra of different plants were examined and compared under
varying illumination to describe how different classes may be
distinguished.
Well-known pure pixels of hyperspectral image were used
directly as reference spectra. In that case, the reflectance of the
pixel must be derived only from material that is defined by the
pixel. If there is variation in reflectance features inside one
class, more reflectance spectra will be needed.
Several methods were used to calculate reference spectra from
the training areas in this study. At first, average values of pixel
values were calculated for every 17 channels. By choosing
median of pixel values was the second method to determine
International
ns
reference Spt
reference. It r
e»
Smaller test |
more detailet
this area (Fig
and it include
Figure 3: 10
p
si
68 T
58
Reflectance value
w
©
Figure 4: Re
:
Besides, bri
reflectance :
other and tl
were not ca
to generali
descriptive
Pure spectri
used as refe
the each p
(Figure 4) v
3.4 Refer:
Comparisoi
spectra wet
images (Fi,
help the in
bright imag
dark pixels
mean that t
