International Archives of the Photogrammetry, Remote Sensing
4. RESULTS
4.1 Feature selection for ERS SAR-images
4.1.1 Individual ERS-images: Figure 1 represents the average
separabilities of land cover classes as function of average
transformed Bhattacharyya-distance. The best separabilities
have been acquired using images taken 5.5.1999, 16.4.1999 and
8.10.1999, but even in these cases the average separability is
rather low. The corresponding weather conditions have been
full snow cover with raining wet snow, 50% snow cover with
raining wet snow quite heavily, and it has been raining
(Pulliainen, 2004). So, it seems that the best conditions in order
to separate these land cover classes are wet snow or ground. The
worst separabilities have been acquired using images 14.7.1999
(no rain) and 27.10.1999 (some rain). Class water is the most
separable class, the worst are agricultural field and open land.
Median filtering of the intensity images increases the
separability. The separabilities are higher with 25m pixel size
than 12.5m pixel size.
Average separability of land cover classes
i
08r 1 ]
9 |
Q 1
S |
1 05! |
©
© |
2 |
E |
= |
S |
= 04 |
|
x |
tea |
9 |
©
E |
Ae |
202 |
v 0? |
c |
S |
= |
|
LE eed emma emma tee ———————— Te ad
0 1 2 3 4 5 6 7 8 9
Image
Figure 1. The average separabilities of land cover classes as
function of average transformed Bhattacharyya-distance, dashed
line means that separabilities have been computed from original
ERS-intensity images, solid median filtered intensity images,
solid line with "x" texture feature Mean and solid line with "o"
texture feature Angular second moment.
In the case of tree species vs. development class, the
separabilities between classes are very low. The best image is
taken 14.7.1999 in dry and warm conditions. In the case of tree
species vs. soil type, the sepa "abilities between classes are very
low. The best image is taken 31.3.1999, in full wet snow cover
and rainy (water) conditions.
4.1.2 Texture features: The separabilities of texture images
varied a lot depending on the used texture feature. The best ones
were Angular Second Moment and Mean, their average
separabilities as function of image are represented in figure |.
The behaviour of the average separability is very similar than in
the case of intensity images. In the case of texture feature
Angular Second Moment, the most separable land cover class is
water, the worst are pine forest and open land. In the case of
texture feature Mean, the most separable land cover class 1s
water, the worst pine, deciduous forest and open land.
As the classification system is tree species vs. development
class, the separabilities are rather low. The most separable
classes are middle aged pine in image taken 8.10.1999 and
spruce sapling in image taken 31.3.1999. As the classification
and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
system is tree species vs. soil type, the separabilities are low.
The most separable classes are pine and spruce on mineral soil.
4.1.3 Best SAR-images: The best subsets of ERS-images were
selected using Branch-and-Bound algorithm with average
interclass divergence as selection criteria. The four most
important images were taken 5.5., 16.4, 8.10. and 9.6.1999, in
the case of intensity images and land cover classes. Two of
these images have been taken in wet snow conditions, one in
rainy and on in rather dry conditions.
The most important texture features varied a lot depending on
the classification system. The three most important features for
land cover classification were Mean, Entropy and Standard
deviation, and the worst was Homogeneity. The three most
important features for tree species vs. development class
classification were Homogeneity, Contrast and Dissimilarity,
and the worst was Angular Second Moment. The three most
important features,for tree species vs. soil type classification
were Angular Second Moment, Mean and Correlation, and the
the worst was Dissimilarity.
4.2 Classification of MODIS-images
Fuzzy means were calculated using weekly NDVI maximum
images and training data for every class. The idea was that the
fuzzy means could be used as training data in a supervised
classification. Figure 2 represents these means for land cover
classes. The beginning of the growing season can be seen
during weeks 16-18 from the beginning of the year. Lower
values during the weeks 23 and 27 are probably due to bad
weather.
— water pins —- Fir
deciduous tree agricultural land peat
— Wamp zoil
se M M -
0,6
0,55
Fuzzy mean
15 16 17 19 20 21 22 23 z4 z5 26 2
Figure 2. The fuzzy means of different land cover classes
computed from MODIS NDVI-mosaics.
The aim of the classification of MODIS NDVI-mosaics was (0
produce the proportions of different land cover classes for each
MODIS pixel. First, the fuzzy means were used as training data
for Spectral Angle Mapper and Spectral Unmixing
classifications. Unfortunately the results were quite poor. Fuzzy
supervised classification was also carried out. After calculating
the fuzzy means and fuzzy covariance matrix, the membership
values for each class were computed. Results of this method
were slightly better than previous ones.
Due to poor results of previous algorithms, the Bayesian
Maximum Likelihood classification was carried out. Training
data pixels whose proportion of the main class was more than
50 94 formed the training set of that class. These pixels were
decided to represent absolute and single classes. The
930
COV
The
and
accu
Fore
agri
is c
class
bene
are
than
imag
that
main
accu
2,1
class
main
open
arour