International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
2.2 Ground truth
The area of Hyytiälä research station is covered by standwise
forest inventory made during 1995 - 2001. Data collection is
based on visual interpretation of aerial photographs and field
measurements. Inventory contains almost 4000 stands, but these
are not all applicable in this study. Stands are rather small; the
average, median and maximum sizes are 2.1, 1.2 and 428.6
hectares, respectively.
2.3 Classification systems
In order to study the information properties of the satellite
images, different kind of classification systems were created
using standwise forest inventory. Training and test sets for
classification were selected using random sampling. The desired
number of samples per class was 1000 but it is smaller for some
small classes. The border pixels between stands were removed
before sampling.
Training data for MODIS-classification was acquired by
estimating the proportions of land cover classes for each
MODIS-pixel. This approach was chosen because the size of
MODIS-pixel is 6.25 ha and the average size of stands is 2.1 ha.
2.3.1 Land cover / use classification: This classification was
used to determine the suitability of the used satellite images to
discriminate general land cover types. There were 2223 stands
in this classification and their average, median, minimum and
maximum sizes were 2.6, 1.6, 0.07 and 428.6, respectively. The
classes in this classification are, including statistics as number
of stands, average size (ha), number of pixels, number of pixels
in training set and number of pixels in test set:
Water: 23, 7.1, 2133, 556, 518
Pine dominated forest: 1382, 2.5, 45367, 1347, 1361
Spruce dominated forest: 558, 2.1, 15314, 1065, 1114
Deciduous tree dominated forest: 88, 1.8, 2304, 573, 531
Agricultural land: 1, 77.9, 1197, 317, 276
Open bog: 36, 14.3, 2031, 520, 498
Open land: 135, 2.5, 4723, 1024, 1022
m)
n9 tA
2.3.2 Tree species vs. development class: This classification
was used to determine the suitability of the used satellite images
to discriminate tree species according to the amount of trees. In
other words, is it better to use these images to separate tree
species or the amount of trees. There were 1652 stands and their
average, median, minimum and maximum sizes were 2.6, 1.8,
0.14 and 63.0, respectively. The classes in this classification
are:
I. Pine, sapling: 275, 2.6, 9015, 1021, 996
2. Pine, young stand: 443, 2.7, 15515, 1075, 1130
3. Pine, middle aged stand: 290, 3.4, 11856, 1079, 1073
4. Pine, regeneration maturity: | 12,2 1, 3073, 973, 365
5. Spruce, sapling: 97, 2.5, 3037, 959, 854
6. Spruce, young stand: 63, 1.9, 1748, 566, 492
7. Spruce, middle aged stand: 162, 2.3, 4823, 1050, 1049
8. Spruce, regereration maturity: 133, 2.3, 4201, 955, 032
9. Deciduous, sapling: 27, 2.2, 898, 245, 203
10. Deciduous, young stand: 35, 1.9, 965, 261, 215
11. Deciduous, middle aged stand: 15, 1.6, 288, 113, 120
2.3.3 Tree species vs. soil type: This classification was used to
determine the suitability of the used satellite images to
discriminate tree species according to soil type. In other words,
does the soil type have some effect to the tree species
classification. There were 1718 stands and their average,
median, minimum and maximum sizes were 2.6, 1.7, 0.07 and
63.0, respectively. The classes in this classification are:
|. Pine on mineral soil: 970, 2.8, 33693, 986, 1023
2. Pine on hardwood swamp : 27, 1.4, 813, 226, 182
3. Pine swamp: 199, 2.6, 6971, 1011, 1044
4. Spruce on mineral soil: 337, 2.3, 10012, 1026, 1017
5. Spruce on hardwood swamp: 112, 2.2. 3801, 1100, 1037
6. Deciduous on mineral soil: 47, 2.1, 1317, 344, 310
7 Deciduous tree on hardwood swamp: 26, 1.6, 728, 251,
197
3. INTERPETATION METHODS
In order to extract relevant information and produce as good
classification as possible, there is need to fuse these two
different kinds of satellite data sets. Data fusion can be
performed on different levels (Pohl and van Genderen, 1998):
e Pixel based fusion: This means that the measurements or
measured physical parameters have been fused. In other
words, the feature vector is combined directly from
different datasources.
e Feature based fusion: This means that features have been
extracted from different data sources using e.g. image
segmentation. In this case the features can be e.g. size,
shape and average intensity level of areas. These features
form feature vectors describing the extracted objects.
e Decision based fusion: This means that the objects have
been identified from individual data sources and then these
interpretation results are combined using e.g. rules to
reinforce common interpretation.
Data fusion in this study is mostly based on pixel based fusion,
but also one kind of decision based fusion is tested. Pixel based
fusion is performed by constructing different featuresets.
There is a large amount of data, so different methods for feature
extraction and selection are needed. Feature selection means
that the best set of images is chosen from all images using some
criteria like the separability of classes. Feature extraction means
that new images are computed from the original ones containing
as much relevant information as possible. The classification of
featuresets using different classification systems was carried out
using Bayes rule. Different classification methods were tested
with MODIS NDVI-mosaics and in the end these were
classified using Maximum Likelihood classifier.
3.1 Feature selection
3.1.1 Bhattacharyya-decision theoretic distance: Class
separability was measured using the Bhattacharyya distance. It
is a probabilistic distance between two classes. Classes are
supposed to be normally distributed, so classes are defined by
their mean vectors and covariance matrices (Devivjer et.al,
1982). Then Bhattacharyya distance was transformed so that the
range of distance would be between 0 and 2, the latter meaning
perfect separability.
3.1.2 Branch-and-bound algorithm: Branch-and-Bound
feature selection algorithm was used to select the best subset of
images from all images according to the separability of classes.
The selection criteria was the average interclass divergence. The
divergence is a measure of separability between two classes,
computed using class means and covariances. It is assumed that
the classes are normally distributed (Devivjer et.al., 1982).
928
Inte
3.2
dec
we
Inc
(NI
gro
mo
of
clo
3.2
def
sub
diff
text
disc
of
ima
(Ha
dim
whi
text
Ang
Enti
3.2.
tran:
COO!
of d
new
COrri
latte
spec
mea:
of in
o
Uo
P
Pixel
featu
spac
]-
La
