1. INTRODUCTION
Artificial neural networks have been used for image pro-
cessing and have shown great potential in classification
of remotely sensed data. However, the amount of data
necessary for training a neural network has not been
addressed. Benediktsson et al. (1990) classified an
image (135 x 131 pixels) using a neural network with
the back-propagation learning algorithm. They trained
with approximately seven percent of the image data and
obtained a training accuracy of 93%. Hepner et al.
(1990) performed a neuro-classification of a four-band
(bands 1, 2, 3 and 4) LANDSAT Thematic Mapper
(TM) image (459 x 368 pixels ) with four land-cover
categories (water, urban, forest and grass). They used
100 (10 x 10) pixels per category for training the neural
network classifier. Two LANDSAT TM images were
enhanced with a digital land-ownership data and then
classified for crop residues (Zhuang et al, 1991;
Zhuang, 1990). The neural network classifiers were
trained with approximately ten percent of the TM data,
and an overall accuracy of more than 90% was obtained
for each classification. From these neuro-classifications,
one to ten percent of image data were used for the train-
ing of the neural networks. Therefore, the amount of
data used for the training needs to be investigated.
The objective of this study was to investigate the
amount of image data necessary for training a neural
network classifier. A LANDSAT TM image was
classified with the classifier, and 5%, 10%, 15%, and
20% of the TM data were used for the training.
2. MATERIALS AND METHODS
2.1 LANDSAT TM Data
The LANDSAT TM scene used in this project was
acquired 29 July 1987. The scene covered an approxi-
mately 10.36 km? area (107 x 107 pixels), including
sections 3, 4, 9, and 10 located in T28N, RSE of Rich-
land township, Miami County, Indiana, U.S.A. Seven
categories of land cover for these sections included
corn, soybeans, forest, pasture, bare soil, and river. The
ground observation data were provided for section 9.
Aerial photographs from 1987 were available for this
study area. The U.S. Geological Survey 1:24,000 topo-
graphic map of the Roann, Indiana Quadrangle was also
used as a reference.
2.2 Neural Network
The neural network used in this study was configured as
a three-layer back-propagation network, including input,
hidden and output layers. Adjacent layers were fully
interconnected. The input layer was composed of an
Nx8 array of binary-coded units, corresponding to N
bands (N = 7 in this study) of the 8-bit LANDSAT TM
data. Twenty units were assigned to the hidden layer,
and six thermometer-coded units in the output layer
referred to the six categories of land cover. With ther-
mometer coding, for example, category 4 of the six
530
categories would be represented as 1 in four most-
significant bits and 0 in the remaining two bits (4-11 1
100).
For the training of a neural network, the TM data were
fed to the input layer and propagated through the hidden
layer to the output layer, and then the differences
between the computed outputs and the desired outputs
were calculated and fed backward to adjust the network
connections (weights). This process continued until the
maximum of the differences was less than or equal to
the desired error. Additional details of the network are
given in Zhuang (1990).
The neural networks simulator was NASA NETS
(Baffes, 1989), which runs on a variety of machines
including workstations and PCs. The simulator provides
a flexible system for manipulating a variety of
configurations of neural networks and uses the learning
algorithm of the generalized delta back-propagation.
The NETS software was run on SUN SPARC worksta-
tions for image classification. Interface routines were
developed to make NETS suitable for image
classification (Zhuang, 1990).
2.3 Neuro-Classifications
The neural network classified an unknown pixel based
on the knowledge learned from a training data set. We
trained a neural network separately with 5%, 10%, 15%,
and 20% TM data. Therefore, four neural networks with
the same configuration were separately trained
corresponding to the various percentages of training
data. These four neural network classifiers were named
NN-5%, NN-10%, NN-15%, and NN-20%. For the study
area, training samples were selected for six land-cover
categories based on the corresponding reference infor-
mation, including the ground observation data, the aerial
photographs, and spectral features from individual
categories. The training data for category river were
obtained by an unsupervised classification (clustering)
of the portion of the image containing the river.
2.4 Normalization of Classification Results
With the iterative proportional fitting procedure, a con-
tingency table can be standardized to have uniform mar-
gins for both rows and columns in order to examine the
association or interaction of the table (Fienberg, 1971).
The classification results were summarized as a confu-
sion matrix for each classifier. Individual entries of the
confusion matrix were divided by the table total, and the
result produced a contingency table. The contingency
table was normalized with the iterative proportional
fitting procedure. The procedure made the row and
column margins consecutively equal one. A standard
function from SAS software (SAS Institute, 1988a) was
used to implement the procedure on contingency tables.
Before implementing the iterative proportional fitting
procedure, we eliminated zero counts in a contingency
table using the method of smoothing with pseudo-counts
(Fienberg and Holland, 1970).
pre
cle
cle
frc
di
an
tra
pr
cy
Ce
we
ne
WC