[Cover] [Contents] [Index] |
Page 146
Table 3.2 Comparison of accuracy of classification using different types of neural network
Network type | (%) Total accuracy | Kappa | Parameter values and network structure |
Multilayer perceptron | 60.28 | 0.521 | Learning rate: 0.2 Structure: 7\28\48\7 |
SOM | 45.83 | 0.376 | Initial learning rate: 0.1 Structure: 7\(7×7) |
SOM | 60.91 | 0.531 | Structure: 7\(12×12) |
SOM | 61.17 | 0.531 | Structure: 7\(25×25) |
Counter-propagation | 48.33 | 0.382 | Learning rate: 0.2 Structure: 7\49\7 |
Counter-propagation | 46.19 | 0.369 | Structure: 7\144\7 |
Counter-propagation | 49.17 | 0.397 | Structure: 7\625\7 |
Fuzzy ARTMAP | 49.39 | 0.394 | Vigilance: ρa=ρb=0.8 Learning rate.: βa=βb=1.0 |
Fuzzy ARTMAP | 57.31 | 0.488 | Vigilance: ρa=ρb=0.97 Learning rate: βa=βb=1.0 |
Fuzzy ARTMAP | 57.01 | 0.487 | Vigilance: ρa=ρb=0.99 Learning rate.: βa=βb=0.2 |
from each type of neural network are shown in Plate 2, and the corresponding confusion matrices are shown in Table 3.3. The multilayer perceptron and the SOM networks (candidates (b) and (c)) achieved the highest classification accuracy of around 60% (with Kappa values of around 0.52), a value that is roughly 3 % higher than the best performing fuzzy ARTMAP, and around 10% better than the counter-propagation network.
Different networks show different reactions to changes in the number of clusters (i.e. neurones in the mapping cortex in SOM, hidden neurones in counter-propagation, and vigilance levels in fuzzy ARTMAP). In the case of the SOM, a network with 7×7 output neurones generates the accuracy of 45.83% (kappa=0.376). If a 12×12 layer of output neurones is used, the accuracy increases considerably to 60.91% (kappa=0.531). However, a network with a layer of 25×25 output neurones showed no significant improvement in classification accuracy, although training time rose considerably, to around 950 CPU minutes for both unsupervised and supervised training. The same behaviour is also exhibited by the fuzzy ARTMAP. When the vigilance level was set to 0.8, a classification accuracy of 49.39% (kappa=0.394) was obtained. When the vigilance value was set to 0.97, accuracy rose to 57.31% (kappa=0.488). A further increase in the vigilance value to 0.99 and a decrease in the learning rate to 0.2 showed no improvement. In contrast, the classification accuracies associated with the three counter-propagation network showed little variation, with output values in the range 46–49%.
[Cover] [Contents] [Index] |