International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
  
  
  
  
  
  
  
0.75~1.0 | 02 200-250 | 0.2 200-250 | 0.2 
Location Criterion | Slope (1) Slope (2) 
(Roads) 
Buffer Score..| SLOPE Score SLOPE Score 
size value value 
« 50m 1.0 <5 0.0 60-75 0.0 
50-100 0.8 5~15 0.09 >75 0.0 
100-150 0.6 15-30 0.52 
150-200 0.4 30-745 0.35 
200-250 0.2 45-60 0.04 
  
  
  
  
  
  
  
  
Table 2 Interpretation Criteria 
2.3 An Artificial Neural Network (ANN) Classifier 
An Artificial Neural Network (ANN) is a simulation of the 
functioning of the human nervous system that produces the 
required response to input (Robert, 1990). ANN is able to 
provide some of the human characteristics of problem-solving 
ability that are difficult to simulate using logical, analytical 
techniques. One of the advantages of using ANN is that it 
doesn’t need a predefined knowledge base. ANN can learn 
associative patterns and approximate the functional relationship 
between a set of input and output. A well-trained ANN, for 
example, may be able to discern, with a high degree of 
consistency, patterns that human experts would miss. In a 
neural network, the fundamental variables are the set of 
connection weights. A network is highly interconnected and 
consists of many neurons that perform parallel computations. 
Each neuron is linked to other neurons with varying coefficients 
of connectivity that represent the weights (sometime is refereed 
as strengths in other literature) of these connections. Learning 
by the network is accomplished by adjusting these weights to 
produce appropriate output through training examples fed to the 
network (Zurada, 1992). 
The multilayer perceptron (MLP) is one of the most widely 
implemented neural network topologies. The article by 
Lippman is probably one of the best references for the 
computational capabilities of MLPs. Generally speaking, for 
static pattern classification, the MLP with two hidden layers is a 
universal pattern classifier. In other words, the discriminant 
functions can take any shape, as required by the input data 
clusters. Moreover, when the weights are properly normalized 
and the output classes are normalized to 0/1, the MLP achieves 
the performance of the maximum a posteriori receiver, which is 
optimal from a classification point of view. In terms of mapping 
abilities, the MLP is believed to be capable of approximating 
arbitrary functions. This has been important in the study of 
nonlinear dynamics, and other function mapping problems. The 
MLPs are trained with error correction learning, which means 
that the desired response for the system must be known, as well 
known as backpropagation algorithm (Zurada, 1992). The 
objective of learning is to minimize the error (RMS in this case) 
between the predicted output and the known output. 
An MLP type neural network model was utilized in this work 
using NeuroSolutions 4.24 software (NeuroDimension, 2004) 
developed by NeuroDimension, Inc. The architecture of a 
network that consists of (a) one input layer that contains 4 input 
variables, (b) one hidden layer of 5 nodes, (c) one output layer 
that contains 1 output variable, and (d) connection weights that 
connect all layers together. 
There are two important parameters including a learning rate 
coefficient (Eta) and a momentum factor (Alpha) during 
training. In general, Eta's valid range is between 0.0 and 1.0. 
Although a higher Eta provides faster learning, it can also lead 
to instability and divergence. A small Eta offers improved 
numerical convergence, however training time is greatly 
increased. When a new ANN training is initiated, the user must 
provide a starting Eta value. It is advisable to start with a small 
number because it is conservative. When a value in the range of 
0.001 to 0.1 is used, it normally starts a smooth training process 
without the risk of divergence. 
The Alpha damps high frequency weight changes and helps 
with overall algorithm stability, while promoting faster learning. 
For most of the networks, Alphas are in the range of 0.8 to 0.9. 
However, there is no definitive rule regarding Alpha. Higher 
momentum values (between 0.8 and 0.9) are most commonly 
used since the damping effect usually helps training 
characteristics. If training problems occur with a given alpha 
value, different values can be tried. In NeuroSolutions, the user 
can define this parameter After several times of test, the alpha 
value is set to be 0.7 in this study. 
The transfer function for PEs serves the purpose of controlling 
the signal strength for its output. The input for the transfer 
function is the dot product of all PEs’ input signals and weight 
vectors of the PE. The four commonly used transfer functions 
are the Sigmoid, Gaussian, Hyperbolic Tangent and Hyperbolic 
Secant. In general, the Sigmoid function {1/(1+e*)} will 
produce the most accurate model, but the learning rate will be 
slower as compared to other functions. The Sigmoid function 
acts as an output gate that can be either opened at 1 or closed at 
0. Since the function is continuous, it allows the gate to be 
opened partially (any value between 0 and 1). Hyperbolic 
Tangent is selected as the transfer function in this study. 
Cross validation is a highly recommended method for stopping 
network training in the NeuroSolutions. This method monitors 
the error on an independent set of data and stops training when 
this error begins to increase. This is considered to be the point 
of best generalization. The testing set is used to test the 
performance of the network. Once the network is trained the 
weights are then frozen, the testing set is fed into the network 
and the network output is compared with the desired output. 
Twenty percentage of training data is used to be a cross 
validation and test dataset in this work. 
3. CASE STUDY AND DISCUSSIONS 
3.1 Test datasets and Pre-processing 
Jiu-fen-ell mountain is selected as the test area, which is a 
typical area of landslides especially after by the big shock of the 
Chi-Chi earthquake at Nantou County of central Taiwan on 
1999/09/21. Datasets collected for this study include Quickbird 
images, digital vector maps including river lines and roads 
obtained from 1:5000 photomaps, DTM, and airborne LIDAR 
data (Shih, 2002). 
Quickbird images are registered to the vector datasets by using 
image-to-map function of ENVI 3.5, which is applying an 
affine transformation as shown in Figure 1, where the false 
color image is a composite of bands NIR, G, and B. Roads are 
designated as yellow colour and river as blue colour. 
NDVI for colour tone criterion is executed using the function 
TRANSFORM>NDVI of ENVI 3.5, as shown in Figure 2. 
Digital Elevation Model (DEM) of the study area is abstracted 
from airborne LIDAR data, retrieved by a Fortran program 
developed by the authors, to match with the satellite image. 
Thus, regular DEM is generated by interpolation of inverse 
distance and used for generating ridge-lines and slope gradients. 
576 
Int 
Th 
Sy! 
La 
IO 
Th 
siz 
fac 
As 
ge 
his 
wi 
co 
thi 
be 
Th 
ne 
int 
nei 
da! 
rar 
for 
Er 
Col 
no 
cy 
wl 
she 
be 
tra 
cla 
Th 
res 
fac 
e 
—— =. NO) PN 0 10] 
-