The International Archives oj the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
191
3.29%. This is due to the fact, that the loss of information from
the urban areas is well captured with the structural features
described in section 2.3. 4
Class
1
Class
2
Class
3
Class
4
Class
5
Class
6
Class
7
Villages
0.55
0.09
0.22
0.05
0.13
0.00
0.00
Mountains
0.10
0.81
0.00
0.00
0.05
0.00
0.02
Fields
0.19
0.05
0.64
0.05
0.18
0.00
0.00
USA
0.06
0.00
0.04
0.82
0.05
0.00
0.02
Europe
0.09
0.05
0.11
0.07
0.60
0.03
0.05
Airports
0.00
0.00
0.00
0.01
0.00
0.97
0.00
Common
0.00
0.00
0.00
0.00
0.00
0.00
0.91
Table 3: Confusion matrix of a SVM linear kernel classification
on 497 images with 7 classes with 30 out of 32 features selected
by FLD.
Class
1
Class
2
Class
3
Class
4
Class
5
Class
6
Class
7
Villages
0.83
0.00
0.15
0.00
0.05
0.02
0.03
Mountai
ns
0.04
0.83
0.01
0.00
0.00
0.00
0.00
Fields
0.04
0.08
0.82
0.01
0.00
0.00
0.01
USA
0.01
0.00
0.00
0.92
0.12
0.02
0.01
Europe
0.08
0.04
0.02
0.07
0.84
0.02
0.02
Airports
0.00
0.05
0.00
0.00
0.00
0.96
0.00
Common
0.00
0.00
0.00
0.00
0.00
0.00
0.93
Table 4: Confusion matrix of a SVM linear kernel classification
on 497 images with 7 classes with 20 out of 36 features selected
by FLD.
4. INDEXING OF LARGE SPOT5 IMAGES
An image is indexed by a set of keywords representing the
content of an image. These keywords are usually limited in
numbers and are dependent on application scenarios.
Classification is often used as a pre-processing step for indexing.
A careful indexing of an image database assists efficient
retrieval of image content. The workflow of our indexing
method is divided into three steps as follows.
4.1 Step 1: The Database
The image database can be viewed as two sets disjointly
partitioned to contain images or segmented images in one set
and features extracted from images in another set. We will
indicate the image set as Si, and the feature set as S F . A pointer
is used between Si and S F to address a image to its associated
feature set. The information extracted in terms of structural
features from the large image archive of 497 images, each of
size 512x512 pixels, categorized into 7 classes are kept in a data
file. The off-line process of this data file creation is done only
once, and in case of a new entry, the information extracted from
this image is augmented with the existing data file. The pointer
is appropriately assigned the address of this new entry. This
will be used as the “training” set later in the classification task.
4.2 Step 2: The Feature File
The off-line process for the user given a large image is as
follows: the large image of size 5120x5120 pixels is
automatically divided into non-overlapping image patches each
of size 512x512 pixels. During this process an association
pointer is asserted from the image patches, defining its spatial
position in the large image. The road network extraction, its
graph representation and the urban area segmentation methods
are applied in parallel on the image set (100 images). The
structural features from the graph representation and the urban
areas are stored in a file. The images are a priori randomly
labelled with classes from 1 to 7. This will be later used as a
“testing” test against the above defined “training” set in the
classification task.
4.3 Step 3: The Classification
In many satellite image classification works, the a priori
information about the class label configuration is available and
it is very essential and crucial to combine this information into
the classification process to obtain a reliable answer. Standard
SVM do not provide any estimation of the classification
confidence and thus do not allow us to comprehend any a priori
information. Probabilistic SVM provides us with a solution as
to construct a classifier to produce a posterior probability
P(class = c|input) which allows us to take a quantitative
decision about the classification (Platt, 1999). In this work we
used a one-vs-rest Gaussian SVM classifier with a=10. The
choice of the Gaussian standard deviation, a, which controls the
width of the kernel is hard to assert in practical situations. In
this study we considered the kernel value which gave us the
least training error.
The results of the probabilistic SVM output can be interpreted
as follows: the classifier output should be a calibrated posterior
probability. First the SVM is trained and then the parameters A
and B of an sigmoid function (see Equation 1) are estimated
from the training set (f„ yj) to map the output of the SVM into
probabilities. The predicted label of an image is the one with
the largest probability value. The large SPOT5, 5m resolution
image, Figure 4(a) of Los Angeles is well classified with a
classification accuracy of about 85%.
The classification image resulting from the probabilistic SVM,
Figure 4(b), shows that certain areas are classified as Europe
urban. This can be explained from the fact that either the
classification probabilities are low or they are comparable with
the neighbouring classes. The other reason for this is the fact
that the network structures in these areas are similar to the one
found in many European urban structures. The superimposed
image in Figure 4(c) validates the classified regions with
ground truths from Figure 4(d).
5. CONCLUSION
Classification of large satellite images with patches of images
extracted from them is a novel idea in the sense that the patches
considered contain significant coverage of a particular type of
geographical environment. Probabilistic SVM provides us with
a quantitative analysis of the classification. This method
provides a basis for more complex analysis of large satellite
images. The effect of overlapping patches on classification is
not reported. This may be an interesting study, as it can help to
better classify the images. Moreover, image patches of different
sizes can also be used to improve the classification performance.
Our indexing method with the above mentioned perspectives
can be adapted with existing and future image information
mining systems for EO archives.