anbul 2004
mn work of
\ccuracy of
Sensing of
d'images
octorat en
ee.
tion, Gibbs
ges. IEEE
>. PAMI-6,
[. HL. and
assification
, 1994, vol.
fication of
ification of
ience and
Multisource
d Imagery
'ensing and
omber 89^
)3-7060-0.
. Study of
contextual
um of the
oratories,
fusion by
lite images.
LL} - 2004
ensing and
tic solution
ial of the
ges. Notes
ince, 95 p.
issification
e study of
981. Pixel
p JEFE.
elligence,
"cognition.
MULTISPECTRAL LANDSAT IMAGES CLASSIFICATION USING A DATA
CLUSTERING ALGORITHM
Yan Wang ® *, Paul Neville, Chandra Bales^, Mo Jamshidi *, Stan Morain b
? Dept. of Electrical and Computer Engineering and Autonomous Control Engineering (ACE) Center, University of
New Mexico, Albuquerque, New Mexico, 87131 — yanwang@unm.edu
? The Earth Data Analysis Center (EDAC), University of New Mexico, Albuquerque, New Mexico, 87131
KEY WORDS: Multispectral, Image, Classification, Networks, Fuzzy Logic, Algorithms
ABSTRACT:
This paper presents a new application of a data-clustering algorithm in Landsat image classification, which improves on
conventional classification methods. Neural networks have been widely used in Landsat image classification because they are
unbiased by data distribution. However, they need long training times for the network to get satisfactory classification accuracy. The
data-clustering algorithm is based on fuzzy inferences using radial basis functions and clustering in input space. It only passes
training data once so it has a short training time. It can also generate fuzzy classification, which is appropriate in the case of mixed,
intermediate or complex cover pattern pixels. This algorithm is applied in the land cover classification of Landsat 7 ETM- over the
Rio Rancho area, New Mexico. It is compared with Back-Propagation Neural Network (BPNN) to illustrate its effectiveness and
concluded that it can get a better classification using shorter training time.
1. INTRODUCTION
Remotely-sensed imagery classification involves the grouping
of image data into a finite number of discrete classes.
Conventionally, statistical Maximum Likelihood Classifier
(MLC), based on normal distribution assumption, is widely
used in remote sensing image classification. However,
geographical phenomena do not occur randomly in nature and
frequently are not displayed in the image data with a normal
distribution. So neural networks with data distribution free have
been applied. The neural network classification depends on
training data and learning algorithms, which cannot be
interpreted by human language or is *a black box". So training
data's selection is important for the neural network
classification. Normally, the training data sets consist of several
thousands of patterns belonging to many (often more than ten)
categories and large volumes of data and the neural network
structure is complicated to adapt to these patterns. So the neural
network training and/or classification time reported are quite
long (Heermann, 1992), ranging in some cases from a few
hours to a few weeks on a conventional computer. Taking also
into account that additional training and classification trials
must usually be performed after selecting a particular neural
network model, its architecture, and its learning parameters, the
need of a methodology for fast neural network training and
classification is evident. Vassilas and Charoun (Vassilas, 1999)
proposed a methodology based on self-organizing maps and
indexing techniques and demonstrated its effectiveness in
classifying multispectral satellite images to land-cover
categories. In this paper, we propose to use a Radial Basis
Function based Clustering (RBFC) algorithm to solve this
problem.
In remote sensing images, a pixel might represent a mixture of
class covers, within-class variability, or other complex surface
cover patterns that cannot be properly described by one class.
These may be caused by the ground characteristics of the
-
* Corresponding author.
17
classes and the image spatial resolution. Since one class cannot
describe these pixels, fuzzy classification has been developed.
In fuzzy classification, a pixel belongs to a class with a
membership degree and the sum of all class degrees is 1. Wang
(1990) modified the MLC algorithm with fuzzy mean and
fuzzy covariance instead of their conventional counterparts.
Foody (1992) embedded the fuzzy concept in all classification
stages, including training, classification and evaluation. RBFC
is based on fuzzy inferences and the fuzzy rules are generated
from the training data. It can also combine human knowledge
in it when it is available. The outputs from RBFC are the
membership degrees of each class.
In this paper the RBFC is provided and applied the land cover
classification of Landsat 7 ETM+ over the Rio Rancho area in
New Mexico. Part 2 gives the RBFC algorithm. Part 3 presents
a study of land cover classification using RBFC and compares
it with Back-Propagation Neural network (BPNN). Part 4
concludes about it.
2. RBFC ALGORITHM
We first describe the mathematical form of the Radial Basis
Function (RBF) rulebase, which is identified by the clustering
algorithm. We will consider a specific case of a rulebase with n
inputs and 1 output. The generalization to m outputs is
described in (Berenji, 1993). The inputs to the rulebase are
assumed to be normalized to fall within the range [0,1]. Each
rule r has the following form, similar to the Takagi-Sugeno-
Kang (TSK) rule:
IF s; is. N (i. 0)...
THEN
and s; is N (&;, 0,)... and s, is N (C, Or)
