Environmental Engineering Reference
In-Depth Information
TABLE 7.3 List of the internal network parameters tested a .
Parameters
Abbreviation
Description
Value
Number of hidden layers
HL
A key factor controlling the topology of neural networks
0
HL
4
Activation function
AF
A linear or non-linear function for processing input data of
neurons.
Log-sigmoid or
Tan-sigmoid
Training threshold
TT
A user-defined threshold determining the contribution of the
input data to the outcome.
0
TT < 1
Learning rate
LR
A user-defined parameter defining the step size of the weight
update
0 . 001 ≤ LR ≤ 0 . 3
Momentum
MO
A user-defined parameter controlling the influence of previous
weight update on current weight update
0 ≤ MO < 1
Number of iterations
IT
A parameter specifying how many times the training algorithm
may iterate toward the targeted training goal
400
IT
2800
a Gradient descent with momentum (GDM) algorithm was used for network training in this focused study.
The entire experiment comprised several major components.
First, we carefully constructed and trained a set of MLP neural
network models with different topologies and training param-
eters. Then, we used these models to classify a satellite image
into several major land cover categories, and we evaluated the
accuracy of each classified map. Based on the classification
accuracies, we further analyzed the sensitivity of these algo-
rithm factors. Second, we compared the classification accuracies
achievedby using the best neural networkmodel and theGaussian
maximum likelihood (GML) classifier. Finally, we summarized
our major findings and recommended several practical guide-
lines when parameterizing the MLP neural networks for image
classification.
3 Exposed land: mainly non-impervious areas with sparse vege-
tation, such as clear-cuts, quarries, barren rock or sand along
river/stream beaches;
4 Cropland/grassland: crop fields and pasture as well as cultured
grasses (such as golf courses, lawns, city parks);
5 Forest: deciduous, coniferous, and mixed forest land; and
6 Water: streams, rivers, lakes, and reservoirs.
7.3.3 Network configuration
and training
We carefully configured 53 MLP neural network models with
different internal parameters combinations. For each neural
network model, the input neurons comprised seven ETM +
image bands (excluding the thermal band due to the coarse
spatial resolution) and the output neurons were six major land
use/cover classes. The general rule is that for the six internal
parameters, only one parameter is allowed to alter at one time
whileholdingtheotherunchanged. In this way, the sensitivity
of neural classification performance with respect to a specific
internal parameter can be assessed. Specifically, to investigate the
impact of hidden layer number, five neural network models were
constructed with the number of hidden layers ranging from 0
to 4 (Table 7.4, Nos 1-5). Then, twenty neural network models
were constructed to address the issues of activation function and
training threshold (Table 7.4, Nos 6-25). Both log-sig function
and tan-sig function were considered, and each was combined
with a set of training threshold values ranging from 0 to 0.9.
Finally, 28 neural network models were configured to assess the
three training parameters, and the range and step of each training
parameter are listed in Table 7.4.
Each of the 53 neural network models was trained with an
identical training sample set that contains 250 pixels for each
land cover class. The training performance was measured by the
root mean square error (RMS). The training goal was set to 0.1
in RMS. The training process stopped when either the maximum
number of iterations or the training goal was reached. Most of the
neural network models successfully converged except the model
with four hidden layers (Table 7.4, No. 5).
7.3.2 Remotely sensed data and land
classification scheme
The remote sensor data used was a Landsat Enhanced Thematic
Mapper Plus (ETM
) image dated on 9 September 1999, which
was georeferenced to the UTMmap projection. The image covers
the northern Atlanta metropolitan area, Georgia, USA. The
landscape in this area is characterized by a mosaic of urban use,
agricultural use, and natural lands, making it an excellent site to
test the effectiveness of different neural network configurations
in image classification. In addition to the ETM + image, we
collected ancillary data through GPS-guided field observations
and the use of high-resolution images available from Google
Earth.
Based on the remote sensor image and the ancillary data, we
designed amixedAnderson Level I/II land use/cover classification
scheme (Anderson et al ., 1976) with the following six major
classes:
+
1 High-density urban use: mostly large commercial and indus-
trial buildings, large transportation facilities, and high-density
residential areas in the city cores;
2 Low-density urban use: mostly single/multiple family houses,
apartment complexes, yards, local roads, and small open
spaces;
Search WWH ::




Custom Search