Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 120

Fig. 5.38 The neural network classification model (Source adapted from www.ndt.net/article/
v05n07/spanner2/spanner2.htm )
A review and analysis of papers published about ANNs before 1994 can be
found in Paola and Schowengerdt ( 1995 ). Example applications of ANNs in
remote sensing image classification for the period between 1994 and 2007 are
given in Schowengerdt ( 2007 ).
The produced back-propagation neural network utilizes the ''generalized delta
rule'' during the learning stage. The network was trained using the same class
training samples which were also used in MLC and SVM. The activation type was
Logistic; the training threshold contribution was 0.90; the training rate was set to
0.10; the momentum rate to 0.90; the training RMS-exit-criteria was 0.10; the
number of hidden layers was 1; and the training cycle (adjustment of weights after
forward and backward propagation of values through the network) was repeated
for a maximum of 1,000 iterations, or until the maximum normalized total error
was less than 0.01, or the maximum individual error was less than 0.001. The last
two situations did not occur, so the training was always performed for 1,000
iterations. (''the individual error is the sum of errors in the output values for one
sample, meaning the difference between target value and output value of each
output node. The normalized total error is calculated as half the sum of the squares
of the individual errors, divided by the number of samples'') (PCI 2001 ). The error
plot was then observed to see whether the value for the normalized total error had
stabilized before the 1,000th iteration. This was the case for all classifications
performed here, although the total error was still between 0.45 and 0.52. In a
second step, the training and momentum rates were lowered to 0.05 and 0.20
respectively, for a slower, more stable training with smaller step increases for an
enhancement of the network weights (PCI 2001 ). 1,000 additional iterations were
improved with these parameters, resulting in final maximum total errors between
0.39 and 0.46.