SIGNIFICANCE-WEIGHTED FEATURE EXTRACTION FROM
HYPER-DIMENSIONAL DATA AND ITS APPLICATIONS
Sadao Fujimura and Senya Kiyasu
Department of Mathematical Engineering and Information Physics
Graduate School of Engineering, University of Tokyo
7-3-1 Hongo, Bunkyo-ku, Tokyo 113, JAPAN
Commission III
KEY WORDS: Algorithms, Feature Extraction, Sensor, Performance, Classification, Accuracy,
Significance Weighted, Hyper Dimensional
ABSTRACT
Extracting significant features is essential for processing and transmission of vast volume of hyper-dimensional data.
Conventional ways of extracting features are not always satisfactory for this kind of data in terms of optimality and
computation time. Here we present a successive feature extraction method designed for significance-weighted supervised
classification. After all the data are orthogonalized and reduced by principal component analysis, a set of appropriate
features for prescribed purpose is extracted as linear combinations of the reduced components. We applied this method
to 411 dimensional hyperspectral data obtained by a ground-based imaging spectrometer. The data were obtained from
tree leaves of five categories, soil, stone and concrete. Features were successively extracted, and they were found to
yield more than several percents higher accuracy for the classification of prescribed classes than a conventional method.
We applied the results of feature extraction for evaluating the performance of current sensors. We used the accuracy of
classification as an index of performance for a specific purpose.
1. INTRODUCTION
Recently the dimension of remotely sensed data becomes
higher and higher because of higher spectral resolution,
increasing number of sensors, and multi-temporal obser-
vations. Airborne Visible Infrared Imaging Spectrometer
(AVIRIS), for example, has 224 spectral bands in the 0.4-
2.5pum region (Vane, 1988). In order to efficiently obtain
necessary information from these hyper-dimensional data,
or in order to transmit the data through a communication
channel, the quantity of data must be reduced. This can
be achieved by extracting significant features.
Here we propose a feature extraction method designed for
significance-weighted supervised classification and present
its application for evaluating the performance of sensors.
The basic idea of our feature extraction is as follows: in
classification of data we have some kind of objectives or
intention. This means that in most of the cases we are
interested in classification of a particular set of classes,
not all of the terrain objects included in the image. Thus
we introduce subjective significance explicitly into feature
extraction. The evaluation to be used is the accuracy for
the particular classes, though conventional feature extrac-
tion methods considering only the average accuracy for
all the classes in the image. The purpose of our feature
extraction is to extract a set of features which optimally
separate one class from another among a particular set of
important classes.
One of the conventional methods of feature extraction
utilizes an exhaustive search for the best subset of sen-
234
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
sor channels using a separability measure between classes
(Swain, 1978), and another uses principal component anal-
ysis (Ready, 1973). The former requires a lot of computa-
tion time to evaluate all the combinations of channels. The
latter is not optimal because the features are not selected
from the viewpoint of discrimination. Canonical analysis
can also be used (Schowengerdt, 1983) and gives better re-
sults than principal components. It extracts the features
which give the best average separability among classes.
However, they are not always suitable for significance-
weighted classification.
Our method extracts appropriate features as linear combi-
nations of orthogonalized and reduced components which
are obtained by principal component analysis. Each fea-
ture is determined successively by considering the distance
from the significant classes until the distance satisfies a
condition.
We applied the results of feature extraction for evaluating
the current sensors’ efficiency for a specific purpose. The
performance of sensors can be evaluated by comparing the
classification accuracy with that by the extracted features.
2. PRINCIPLE OF FEATURE EXTRACTION
2.1 Description of Data
First of all and as usual, we assume that we can get train-
ing data for almost all the classes in an image to derive
feature with: that is, we can estimate the characteristics
of most c
We deno!
vector y
that they
y can be
dispersion
(see Fig.
the covar
respectiv
class cov
that the
assumpti
generality
2.2 Feat
Here, for
two most
all the ot
In genera
bility" of
the perfoi
tures whi
of classes
Our met]
cessing: |
In the pi
E yn)
compone:
z=A'y
dispersioi
ellipsoida
they are
makes th
measured
ity in the
In the se
(Kiyasu,1
which has
*We us
separabili
