taken
atures
red to
bined
‚visual
ecked
S plan
er air
O be
1 for
a user
terrain
and a
; MDP
, town
ottage
lustrial
AN EFFICIENT SUPERVISED CLASSIFICATION METHOD OF REMOTELY SENSED MULTISPECTRAL IMAGES
Hiroshi Hanaizumi, Kentaro Saito, Shinji Chino
College of Engineering, Hosei University
3-7-2 Kajino-cho, Koganei, Tokyo 184, Japan
Phone +81-423-87-6354
Fax +81-423-87-6126
E-mail hana@hana.is.hosei.ac.jp (H. Hanaizumi)
and
Sadao Fujimura
Graduate School of Engineering, The University of Tokyo
7-3-1 Bunkyo-ku, Tokyo 113, Japan
Phone +81-3-812-2111 ext.6900 Fax +81-3-5689-7354
E-mail fuji@k2.t.u-tokyo.ac.jp
Commission IV, Working Group 1
KEY WORDS: Linear discriminant function algorithm, Clustering, Data projection, Histogram, Valley, Binary division tree,
within-sum-of-squares, Training data
ABSTRACT
A method is proposed for supervised classification of remotely sensed multispectral images with high accuracy and high
efficiency. The method Modified Linear Discriminant Function (MLDF) produces a binary division tree by dividing training
data until all the terminal nodes of the tree have only one kind of category. After production of the tree, whole image data are
classified. Numerical simulation indicates the method has as high accuracy as Maximum Likelihood method does and as
high efficiency as a Binary Division Tree classifier does.
INTRODUCTION
Supervised classification is one of the basic processes in
the application of remote sensing technology to various fields.
Itis indispensable analysis of remotely sensed multispectral
images, for example, in environmental monitoring. For
monitoring earth environment using satellite, it is important
that a classifier must have high efficiency as well as high
accuracy. Among the methods proposed so far, maximum
likelihood classifier (MLH) is well-known and most accurate
on the assumption that the generality of training data is
satisfied (Fujimura, 1978), but it requires much time to classify
the image. On the other hand, Binary Decision Tree (BDT)
classifier (Inamura, 1979) is one of the most efficient
methods, but it has lower accuracy than MLH does on the
assumption above. A method having high efficiency as well
as high accuracy has been required.
Here, we propose a new method having both high accuracy
and high efficiency. We call the method as Modified Linear
Discriminant Function (MLDF). The method MLDF is
expanded from BDT. We introduce linear discriminant
function into boundary selection in BDT. The division
boundary of MLDF is determined using binary division
technique in a clustering method BDC-LDF (Hanaizumi,
1995a, b). The boundary is selected among valleys in
density histogram obtained from image data projected onto
327
a single dimensional subspace. As we do not use the
statistics (such as mean and variance) of training data, we
can regard MLDF as a nonparametric classifier.
In this paper, we describe the principle and the procedures
of MLDF. The validity of MLDF is confirmed by numerical
simulation and classification of real remote sensing images.
PRINCIPLE
In the feature space, MLH produces hyper-quadratic
boundaries which theoretically achieve the highest accuracy
with much time for the classification. BDT achieves high
speed with some loss of accuracy by limiting number of
boundaries for binary division of data. It is known that the
accuracy of linear discriminant function (LDF) algorithm is
identical to that of MLH when all training data sets have the
same variance-covariance matrix. By using LDF algorithm
hierarchically, we achieve highly accurate (as well as MLH)
and highly efficient (as well as BDT) .
The basic ideas of the proposed method MLDF are to label
pixels in training area with category identification, to merge
all the pixel data in all training areas into one group and to
apply binary division process to the group so that all data in
a terminal node have the same identification. After
production of decision tree, all image data are classified as
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
