Environmental Engineering Reference
In-Depth Information
TABLE 7.4 ( continued ).
No.
HL
AF
TT
LR
MO
IT
Classification Accuracy (%)
42
1
log-sig
0
0.01
0.7
1000
83.33
43
1
log-sig
0
0.01
0.8
1000
84.00
44
1
log-sig
0
0.01
0.9
1000
82.33
45
1
log-sig
0
0.01
0.8
400
83.00
46
1
log-sig
0
0.01
0.8
700
83.33
47
1
log-sig
0
0.01
0.8
1000
84.00
48
1
log-sig
0
0.01
0.8
1300
84.67
49
1
log-sig
0
0.01
0.8
1600
84.00
50
1
log-sig
0
0.01
0.8
1900
83.33
51
1
log-sig
0
0.01
0.8
2200
81.67
52
1
log-sig
0
0.01
0.8
2500
81.00
53
1
log-sig
0
0.01
0.8
2800
80.33
Note that model 5 failed to converge during the training phase.
HL, number of hidden layers; AF, activation function; TT, training threshold; LR, learning rate; MO, momentum; and IT, number of iterations.
7.3.4 Image classification and
accuracy assessment
deviation. Among these models, the one with single hidden layer
produced the best overall classification accuracy, followed by the
model with zero hidden layer, with two hidden layers, and with
three hidden layers. This finding concurs with those from Shupe
and Marsh (2004) and Kanellopoulos and Wilkinson (1997).
Theoretically, neural network models with more hidden layers
can deal with more complex problems but require a large sample
size to train. When input and output neurons are limited in
number and training size is moderate or relatively small, neural
network models equipped with more hidden layers can become
less effective or even fail to converge in the training phase as
they may end with local minima or overly fit the training data.
Thus, the selection of an appropriate hidden layer number should
consider the complexity of input and output neurons as well as
the training sample size.
Each trained neural network model was used to classify the
ETM + image into the six land cover classes and hence a total
of 52 land cover maps were produced with the exception of
Model 5 that the training process failed to converge (Table 7.4,
No. 5). The classification accuracy of each map was assessed by
using the confusion matrix method that is based on the use of
a reference dataset (Congalton, 1991). It computes the overall
accuracy, user's accuracies, producer's accuracies, and the Kappa
statistic through the comparison of the predicted values and
the actual values of the reference samples. To correctly perform
the accuracy assessment, a reference dataset was collected by
using the stratified random sampling scheme with approximately
50 samples for each land cover class. For each of the 52 land
use/cover maps, a confusion matrix was created, and the overall
classification accuracies were used to evaluate the performance
of each neural network model for land cover classification.
For comparison purposes, the GML classifier was also trained
with the identical training samples and then used to produce a
land use/cover map from the ETM
7.3.5.2 Activation function, training
threshold, and classification accuracy
Table 7.4 (Nos 6-25) and Fig. 7.3(B) suggest that the activa-
tion function type can greatly affect the performance of neural
network models. Clearly, the models equipped with the log-sig
function substantially outperformed the ones with the tan-sig
function, as indicated by the average overall accuracies. When
incorporating training threshold in the comparison, we found
that the models with the log-sig function were less sensitive to
the training threshold values used when comparing to the ones
equipped with the tan-sig function, as indicated by their stan-
dard deviations (2.65 vs. 5.79). For the former group of models,
the best classification accuracy was obtained when the training
threshold value was set as 0. Although the variation of classifi-
cation accuracies by this group of models was relatively small,
a decline trend emerged when the training threshold value was
raised to 0.9. This suggests that a smaller training threshold value
should be used for this group of models equipped with the log-sig
function. For the models with the tan-sig function, the variation
of their classification accuracies was larger, and relatively higher
image. The classification
accuracy was assessed with the same reference samples. We
further compared the classification accuracies achieved by using
the best neural network model (Table 7.4, No. 48) and the GML
classifier.
+
7.3.5 Interpretation and analysis
7.3.5.1 Hidden layer number and
classification accuracy
Based on Table 7.4 (Nos 1-5) and Fig. 7.3(A), it is found that
the performance of neural networks was quite sensitive to the
hidden layer number, as indicated by the relatively high standard
Search WWH ::




Custom Search