Database Reference
In-Depth Information
8.4 Experimental Results
Our experiment uses the SPOT spectral satellite images provided by the Earth
Resource Satellite Receiving Station in Taiwan. These images contain five classes:
soil, city 1, city 2, sea, and forest. Our objective is to classify these classes from
the three-channel satellite images. The results show that when we try to use all
inputs without any dimension reduction to perform the training of a classifier, ei-
ther the SONFIN or the backpropagation (BP) network, the result is poor and the
MSE is very large. We first apply our CNFM to reduce the dimension of inputs
and then perform the training of the classifier.
An illustrative example of classification of a channel is shown in Fig. 8.11. It
gives the details of our proposed system. First we obtain the gray value of a cen-
ter-pixel. Its value is 186. Among those points, we calculate the ASM, CON, IDM,
entropies, and energies from the co-occurrence matrices and wavelet transforma-
tions with a neighborhood size
N = 7
. After we have acquired the input features,
we pass each group of the features into Kohonen's SOM. Each group of input fea-
tures with high dimension is reduced to 2D coordinates. Finally, we use these 2D
coordinates, combined with the coordinates from other channels, as the inputs to
our neural fuzzy network, SONFIN. The result we obtained is the desired class;
here it is class 3.
Tables 8.1, 8.2, and 8.3 show the results of the classification using the pro-
posed CNFM. Here we compared the accuracy among different types of inputs.
The results in Table 8.1 are the best, and we obtain good results because all infor-
mation was used as inputs to CNFM. Table 8.2 shows the result of classification
when gray level information and statistical features are used. This is the general
architecture that most people use as their framework. We can see that the accuracy
decreased in each class. The results in Table 8.3 are obtained using any three gray
values. It can be seen that without sufficient information, good classification is
hard to achieve. Figure 8.12 shows the original satellite image with ground truth
and the classified image using CNFM with three types of inputs. From Fig. 8.12,
we can conclude that CNFM can perform an accurate classification.
Table 8.4 shows the classification results of the
k
-nearest-neighborhood (KNN)
method, BP network, and the CNFM. The inputs of the BP network and CNFM
are first filtered by Kohonen's SOM, whereas KNN uses all features directly. It
can be easily concluded that the KNN and BP network are insufficient to obtain
accurate results. Figure 8.13 shows the MSE (mean square error) of the BP net-
work against SONFIN in our CNFM. We can conclude that the BP network is not
adequate to reduce the MSE due to the complexity of the images, and our system
can tackle this problem successfully. The used BP network has two hidden layers
with 30 hidden nodes in each layer.
Search WWH ::
Custom Search