tof 
cally 
each 
orm, 
two 
0). a 
siven 
esent 
, and 
| and 
ected 
ag all 
id its 
s, and 
This 
onger 
up of 
19 the 
iginal 
These 
] over 
ce the 
ors in 
| input 
n, this 
M. 
ing by 
(CM). 
erated 
ask of 
in the 
vill be 
n. The 
KCM has some important characteristics that enable these 
tasks: 
e As the SOM performs a clustering on the training patterns, 
it’s possible to visualize the spectral classes present in the 
original image through the clusters obtained in the KCM. 
e The SOM property of preserving the topological relationships 
among the input data vectors is reflected in the KCM property 
of preserving these relationships among the clusters so 
obtained, in terms of distances among them. Clusters that are 
close to each other in KCM represent land cover classes which 
posses similar spectral features; 
e The SOM property of preserving the probability distributions 
found in the input data can be verified in the KCM, where 
higher frequency spectral classes in the input data will be 
mapped onto bigger regions in the KCM. 
Therefore, the spectral classes and their samples, which will be 
used in the classification phase, are selected from the KCM 
and not directly from the original image as is usually done. 
2.3 Parallel Implementation of SOM 
The inherent parallelism of ANN is well known. Efficient 
parallel implementation of neural networks both in hardware 
and in software is an active research field. 
In this module for feature extraction the parallel 
implementation of SOM was realized by software, aiming at 
improving the performance in terms of the training time of 
SOM, using a tool developed by the Oak Ridge National 
Laboratory, the Parallel Virtual Machine (PVM). 
PVM is a software system that enables a collection of 
heterogeneous computers to be used as a coherent and flexible 
concurrent computational resource. The individual computers 
may be shared- or local-memory multiprocessors, vector 
supercomputers, specialized graphics engines, or scalar 
workstations, that may be interconnected by a variety of 
networks, such as Ethernet, FDDI. s 
User programs written in C, C++ or Fortran access PVM 
through library routines. Daemon programs provide 
communication and process control between computers. 
For SOM, the basic idea for paralleling the training algorithm 
was to allocate sefs of neurons from the SOM to the 
processors, distributing the training patterns among them, so as 
to reduce the global computational time. 
A comparison between the performance of the sequential and 
the parallel training algorithm of SOM is shown in section 4. 
3. MLP FOR CLASSIFICATION 
Having selected the desired classes and their correspondent 
samples from KCM, the objective of the second phase in our 
proposed system is to perform the final classification of the 
image using a Multilayer Perceptron (MLP) network. 
119 
MLP belongs to the class of feedforward neural networks, 
consisting of a number of neurons which are connected by 
weighted links. The units are organized in several layers, 
namely an input layer, one or more hidden layers, and an 
output layer. The input layer receives an external external 
activation vector, and passes it via weighted connections to the 
units in the first hidden layer. These compute their activations 
and pass them to neurons in succeeding layers. 
The training of the MLP network in performed in a supervised 
way, where the objective is to tune the weights in the network 
such that the network performs a desired mapping of input to 
output activations. 
The MLP network in our system has one hidden layer (fig.3). 
The number of neurons per layer varies according to the 
number of classes and to the size of selected samples to 
perform the training. 
  
Output layer 
Hidden layer 
   
Input layer 
  
  
  
Figure 3: MLP network. 
3.1 Training algorithm for MLP 
Several adaptive learning algorithms for MLP neural networks 
have recently been discovered. Many of these algorithms are 
based on the gradient descent algorithm well know in 
optimization theory. They usually have poor convergence rate 
and depend on parameters which have to be specified by the 
user, as no theoretical basis for choosing them exists. The 
values of these parameters are often crucial for the success of 
the algorithm. An example is the standard backpropagation 
algorithm (Rumelhart et al 1986), which often behaves very 
badly on large-scale problems and whose success depends on 
user dependent parameters like learning rate and momentum 
constant (Moller 1993), which is often the case with RS 
applications, that normally handle large and full of details 
images. 
In this search for alternatives to the poor performance of 
standard backpropagation, normally pointed out as the main 
drawback to a broader utilization in RS image classification, 
this work used an advanced training algorithm for the MLP 
network. 
The MLP learning algorithm used here is an improved version 
of the Scaled Conjugate Gradient (SCG) algorithm presented 
in (Moller 1993). 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996