2.2 Ground truth map
To assess the capability of IRS-P6-LISS IV images for forest
density mapping, an accurate ground truth was prepared
through fieldwork, since there is no update aerial photo or other
very high resolution images. An inventory grid of 500mx500m
and 34 plots (each plot 1 ha) with random systematic
distribution were designed on a topographic map. Coordinates
center of the samples were entered to a handheld GPS. After
revealing the plots in forest, their densities were estimated
qualitatively in density classes of 0-594, 5-1094, 10-15%, 15-
20% and > 20%. There was no forest area with density more
than 25% in the study region. Due to coppice forest and low
BHD of trees, number of trees in hectare could not be an
appropriate parameter to estimate the density. Therefore,
percent of tree canopy cover was estimated in each sample plot.
The resulted density vector map was converted to raster format
(Figure 3). Finally a ground truth map with 3 classes was
produced to be compared with map resulted of satellite image
analysis.
3000 Matars
Figure 3. Sample ground truth overlied on the satellite image of
the study area
3. DATA
3.1 Satellite data and geometric correction
A subset of a map oriented IRS-P6-LISS IV with 3 bands (B3,
B3 and B4) and 10m resolution dated 3 1-August 2007 has been
used. The image underwent level 1G processing (geometrically
and radiometrically corrected) and had no cloud cover. The
subset was rectified to another precise orthorectified IRS-P6
from the same region and year with GCP method (RMSe<5m).
The nearest neighbour resampling method was performed to
produce image with the same resolution of ground truth (5m).
The image was geocoded to the UTM coordinate system.
3.2 Image processing
In order to extract more accurate information from satellite
data, various suitable enhancements such as principal
component analysis [Eastman, 2006] and band rationing were
performed. Since the canopy cover is very low, distance-based
vegetation indices were also calculated to reduce influence of
International Archives of the Photogrammetry, Remote Sensin
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
g and Spatial Information Sciences, Volume XXXIX-B8, 2012
soil background using soil line parameters [Alavipanah, 2003].
These parameters were estimated based on regression analysis
of Red and NIR reflection of different soil types existed in the
study area.
3.3 Image classification
Supervised classification methods were used for image analysis.
The training set into 5 density classes was defined.
The best spectral band-sets were selected using bhattacharrya
distance and transformed divergence criteria based on training
areas. Classification utilizing original and synthetic bands with
maximum likelihood (ML), minimum distance to mean (MD)
and fuzzy classifier was performed. Since the primary results
indicated spectral interference between some density classes,
these classes were merged together and the classifications were
repeated. In order to eliminate single pixels deviating from the
neighbourhood, a majority filter (7x7 pixels = 35mx35m) was
done on the resulted maps. Accuracy assessment of
classification outputs was accomplished through the use of error
matrices detailing producer, user and overall accuracy and an
overall kappa statistic [Congalton & Green, 1999].
4. RESULTS
-Desired coincidence between crest and valley layers of digital
topographic maps and the rectified satellite image indicated
high precisian of the image rectification.
-The ground truth map included of five density classes (0-594,
5-10%, 10-15%, 15-20% and > 20%) was prepared for about
7% of the study area (Figure 3).
-Classification outputs with 5 density classes showed
undesirable overall accuracy and kappa coefficient of 50% and
0.31 respectively. Merging some classes with spectral similarity
improved the result. The best result of forest density
classification was acquired by fuzzy classifier with 3 classes (0-
5%, 5-20%, >20%). The overall accuracy and kappa coefficient
were 70% and 0.44, respectively (Table 1). Figure 4 presents
the result of recent classification.
4051000 |
34051000
4043000 4043000
404000
274000 272000 275000 274000 DE
9 1000 $ 2006 3000 Metsrs
Figure 4. Map of canopy density with three classes resulted
from fuzzy classifier (UTM Zone 41, WGS84)
Int:
T:
Base
class
accu
class
Cons
to ]
class
cont
[Ski
meth
dens
area.
Ove
appr
Higl
aeris
class
In th
perl
Alay
eart