! 2004
7 and
good
e two
an be
3):
nts or
other
from
e been
image
a SIZE,
satures
s have
1 these
les to
fusion,
| based
feature
means
g some
means
taining
tion of
ied out
tested
| were
Class
nce. It
ses are
ined by
r etal,
that the
neaning
-Bound
1bset of
classes.
ce. The
classes,
ned that
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
3.2 Feature extraction
3.2.1 Weekly NDVI-mosaics from MODIS: In order to
decrease the amount of MODIS-data it was decided to compute
weekly NDVI-mosaics. Normalized Difference Vegetation
Indices were computed using red and near-infrared channels as
(NIR-RED)(NIR+RED). Then these NDVI-images were
grouped according to their acquisition week. Finally, the weekly
mosaic was constructed by selecting the maximum NDVI-value
of individual images as mosaic value in order to get rid of
clouds.
3.2.2 Texture features from ERS-images: Texture can be
defined as a variation of the pixel intensities in image
subregion. The assumption is that the intensity variation of
different land-use classes are different and by characterizing
texture by using some measure we can help class
discrimination. Texture features describing the spatial variation
of image grey levels were computed from original intensity
images (12.5 m pixel size) using Haralick's co-occurence matrix
(Haralick etal, 1973). A co-occurrence matrix is a two-
dimensional histogram of grey levels for a pair of image pixels
which are separated by a fixed spatial relationship. Following
texture measures were computed from co-occurrence matrix:
Angular Second Moment, Contrast, Correlation, Dissimilarity,
Entropy, Homogeneity, Mean and Standard Deviation.
3.2.3 Principal component analysis: Principal component
transformation is a linear transformation which rotates the
coordinate axis of the feature space according to the covariance
of data (Richards, 1993). The result of the transformation is a
new set of images, where in principle, the first images
correspond to the information needed in classification and the
latter images correspond to the random components like
speckle. It should be noted that the image variance is used as a
measure of image information and it can depend on the scaling
of images.
3.3 Computed featuresets
Pixel based data fusion was performed by constructing different
featuresets for classification. The selected dimension of feature
space was six. These featuresets were:
l. The six best median filtered ERS-intensity images chosen
from all ERS-intensity images using Branch-and-Bound
algorithm. The size of filtering window was 3x3 pixels.
The chosen images were taken 31.3., 16.4., 5.5., 9.6., 14.7.,
8.10.1999,
2. The principal component analysis was performed to all
. median filtered ERS-intensity images. The six first
principal component images were chosen.
3. The three first PCA-images were computed from median
fillered intensity images. Texture images were computed
using features Mean and Angular Second Moment for all
unfiltered intensity images (12.5 m pixel size), averaging
them to 25 m pixel size, normalizing features to zero mean
and unit variance and performing the principal component
analysis. Three first principal component images were
chosen.
4. The three first PCA-images were computed from median
filtered intensity images. The two first PCA-images were
computed from texture features as previously. MODIS
NDVI-mosaic (week 31) was selected as the sixth feature.
5. The two first PCA-images were computed from median
filtered intensity images. The two first PCA-images were
computed from texture features. The principal component
929
analysis was also performed for all MODIS NDVI-mosaics
and the two first principal component images were chosen.
6. The two first PCA-images were computed from median
filtered intensity images. The two first PCA-images were
computed from texture features. Two features were
computed from the a'posteriori probabilities of Maximum
Likelihood classification of MODIS NDVI-images. The
first MODIS NDVI-feature was the a'posteriori probability
of forest classes. The second feature was the sum of
a'posteriori probabilities of classes agricultural land and
open land.
3.4 Classification algorithms
3.4.1 Bayes rule: Classifications of featuresets were performed
using Bayes rule for minimum error with k-nearest neighbor
density function estimation method (Devivjer etal., 1982).
Number of nearest neighbors, k, varied from 1 to 15. A'priori
probabilities for classes were equal or a'posteriori probabilities
of MODIS NDVI-classification were used as a'priori
probabilities. This is one way to perform decision based fusion;
use the result of low-level interpretation as input to a higher
level interpretation (Schneider et.al., 2003). Classification errors
were estimated using resubstitution and holdout methods,
meaning that the ground truth data was divided to training and
test sets. In resubstitution method the same set is used as
training and test set (optimistically biased method) and in
holdout’ method data is divided to training and test sets
(pessimistically biased) (Devivjer et.al., 1982).
3.4.2 Classification of MODIS-images: The aim of the
classification of MODIS NDVI-mosaics was to produce the
proportions of different land cover classes for each MODIS
pixel. The classifications were made using Spectral Angle
Mapper (Kruse et.al., 1993), Spectral Unmixing (Kruse et.al.,
1997), fuzzy Maximum Likelihood (Wang, 1990) and
traditional Maximum Likelihood (Lillesand and Kiefer, 1994)
classifiers.
3.5 Error measures
The success of classification was measured using error matrix in
the case of featuresets and computing bias, RMSE and
correlation in the case of MODIS NDVI-classification.
3.5.1 Error matrix and measures: One of the most common
means to examine the classification result is to form
classification error matrix which compares the relationship
between reference data and classification result on class-by-
class basis. The columns of error matrix correspond to the
reference data, showing into which classes the reference pixels
have been classified. The rows of error matrix correspond to
classes in the classification result. Several accuracy measures
like Overall accuracy, Producer’s accuracies of individual
classes, User’s accuracies of individual classes and Kappa
coefficient were computed from error matrix (Lillesand and
Kiefer, 1994).
3.5.2 Error measures for MODIS-classification: The result of
the classification MODIS NDVI-mosaic was the proportions of
the land cover classes within MODIS-pixels. In this case the
accuracy of classification was evaluated by computing bias,
root-mean-square-error and correlation between training data
and estimated proportions.
