International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
The received image was subjected to radiometric correction
using VTT’s in-house software (smac_corr.exe) using the
coefficients from Space Imaging (2004).
Figure 1 Ikonos true color image from Suonenjoki, acquisition
date 5.9.2003. The test image area depicted with red square
Figure 2 Ikonos False color image of Suonenjoki test area (R —
NIR chn., G = green chn., B — blue chn., 1 = intensity layer =
PAN chn.)
2.2 Ground data
The Suonenjoki Research Station of the Finnish Forest
Research Institute has measured 327 sample plots for 7
rariables for different tree species in summer 2001. The
Suonenjoki ground data was thoroughly examined visually to
exclude ground sample plots that distort or cause error to the
estimation process.
A total of 74 data points were removed from the original ground
data set (327 points) after a careful visual inspection. These
data points were considered as erroneous or were too close to
borders of very different ground segments. A set of 43 points of
zero (39 points) or near zero (4 points) data were added on
areas regarded as clear cuts or fields with zero total stem
volume. In the resulting ground data set there were thus 296
sample plots (see Figure 3).
m3/ha
Figure 3 Ground data points after data exploration. Plot color
proportional to stem volume (dark/blue V = 0 m3/ha,bright/ red
V = 435 m3/ha)
3. METHODS
FOREST VARIABLE ESTIMATION
3.1 Feature extraction
The spectral features averaged from the radiometrically
corrected Ikonos channels together with the contextual features
calculated from the Ikonos PAN-chromatic channel form the
input feature set to the Forestime estimation process.
The test feature set contained five Haralick features (Haralick,
et. al. 1973): contrast, entropy, inverse difference moment,
homogeneity, and sum average. The associated grey-level co-
occurrence matrix has been calculated for a 15 x 15 pixel
window, and with distance relation of one pixel in both image
directions. The occurrences in the four possible pixel separation
combinations were summed together i.e. the direction
information is lost. The co-occurrence matrix was calculated for
a compressed image of 16 grey levels.
In addition to the Haralick features, a set of four Gabor features
were calculated from the PAN-chromatic channel of the Ikonos
image, using a bank of even-symmetric real-valued Gabor filter
masks (Jain & F. Farrokhnia, 1991). In the method a set of
Gabor filters, covering the image spatial-frequency domain
nearly uniformly, are generated, and a filtered multi-channel
image is produced from the input image.
The tree location tool is determines tree crown locations using
the local maximum filtering (LM filtering) technique on the
Ikonos PAN-chromatic channel (Wulder, et.al, 2002). The
Forestime averaging tool then produces a segmentwise local
maximum density feature as its output. The tree species
proportions tool registers the tree reflectance from the Ikonos
image spectral channels at the locations given by the tree
location tool. Near-infrared reflectance of the nearest pixel of
each located stem is used to determine whether the tree is
broad-leaved or a conifer. If near-infrared reflectance exceeds
the given threshold the tree is labelled as broad-leaved,
336
Internat
————
otherwis
produce:
of broad
3.2 Fes
To get 2
models,
mutual «
creating
SPSS st
SPSS c
marginal
very sm;
increase
with Fo
concentr
variable
containe
features
farly s
augment
estimate
3.3 Seg
The basi
segment:
in the set
relevant
based or
texture Ir
include o
The mu
develope
segmenta
(1980). 7
original i
belonging
method b
34 Vari
The segn
averages
images ar
of input
algorithm
cluster is
target var
this cluste
input sam
2001). Th
in vector
depicted i
