You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

The 3rd ISPRS Workshop on Dynamic and Multi-Dimensional GIS & the 10th Annual Conference of CPGIS on Geoinformatics
Chen, Jun

ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001
1 2 1
Dongmin HUO , Jingxiong ZHANG , Jiabing SUN
(1 School of Remote Sensing and Informatics Engineering, Wuhan University, Wuhan 430079,China)
(2 National Key Lab of Information Engineering for Surveying and Mapping and Remote Sensing,
Wuhan University, Wuhan 430079, P. R. China
KEY WORDS: fully fuzzy; remotely sensed imagery; supervised classification; neural networks; land cover
This paper presents a fully fuzzy strategy for image classification based on a neural network architecture, which promise to overcome
some problems encountered in a conventional crisp even fuzzy classification process. The proposed method was successfully applied in
a classification of land cover with results confirming the flexibility and practicality of this fully fuzzy approach.
The technology of RS (Remote Sensing) is widely applied
to the resource management and environment monitoring. And
the image classification is a very important application aspect.
Generally speaking, the classification of remotely sensed image
is performed by means of visual interpretation or computer
automation or semi-automation. Firstly, the procedure of
classification needs an appropriate classification category
system according to the requirement and ground fact. Next,
every pixel is classified into a certain category by choosing a
classification method. Finally, the average classification
accuracy is calculated by forming an error matrix according to
the other group of independent reference data. The prerequisite
of this conventional crisp classification is that every pixel of the
image and the reference data used for estimating the
classification accuracy are supposed to have a unique category
affiliation. That is to say, suppose the real world can be clearly
defined and measured. In fact, the real world itself is a huge and
complicated multi-dimension system. Among it, so many natural
or semi-natural phenomena can’t easily be classified into some
certain categories. Although the spatial resolution of Remotely
Sensed image is being improved, there are still many pixels
whose spectrum characteristic is the combination result of
response from their neighbors. Besides, the phenomena which
are ‘the same spectrum from different features' or 'the same
features have different spectrum’ are still widely existing. They
lead to the serious limitation of the traditional classification
method. Therefore, this problem of classification method cannot
be resolved only on the basis of statistical analysis.
The theory of fuzzy set is a big hit in the field of RS
classification at present. According to the theory of fuzzy set, an
element or an object, such as a pixel of the RS image, has the
fuzzy affiliation relationship. Many experiments prove that the
fuzzy method can express the ambiguous phenomena and their
characteristics more efficiently than the conventional
classification method does. But the fuzzy classification method
ignores the existing ambiguity of the training process in the
supervised classification. It implies that the training sample is
supposed to be made up of a group of representative objects or
pixels, which can be clearly defined and classified. This actually
follows the conventional method. So it is not the genuine fuzzy
This paper analyses a so-called method of fully fuzzy
supervised classification by means of the brief mathematical
description about the classification procedure. Utilizing the
neural network which has a good mechanism of learning and
generalization and the capability of anti-variance and easily
expanding to a dynamic system, this paper designs a fully fuzzy
classification system. And the system is applied to the land cover
classification of the imagery (Landsat TM) in the study area of
the Delta of the Yellow River in China. This paper also provides
the accuracy evaluation of land cover classification. Its efficiency
can be proved.
Generally speaking, every pixel in the unclassified image is
an object, which belongs to an element or a class of the
category set {1,2,•••, such as spectral brightness and texture character, forming an
eigenvector X . Supposed that the probability of this pixel
belonging to class /(/ G (1,2,• • *,c}) is 7l l and the
probability distributing density of the eigenvector X belonging