Environmental Engineering Reference
In-Depth Information
TABLE 9.7 Overall accuracy, producer's accuracy, user's accuracy, and Kappa coefficient produced by the nearest neighbor classifier
with mean value of the original bands and three PCA bands.
Referece Data
Total
Producer's
User's
B
I
T
G
P
S
Sh
Pixels
Accuracy
Accuracy
B
31
0
0
0
0
3
0
34
93.94%
91.18%
I
1
32
0
1
0
0
2
36
80.00%
88.89%
T
0
0
27
1
0
0
0
28
93.10%
96.43%
G
0
0
2
22
0
0
0
24
91.67%
91.67%
P
0
5
0
0
15
0
2
22
93.75%
68.18%
S
1
2
0
0
0
26
3
32
89.66%
81.25%
Sh
0
1
0
0
1
0
22
24
75.86%
91.67%
Total Pixels
33
40
29
24
16
29
29
200
Overall Accuracy = 87.50%
Kappa Coeficient = 0.85%
TABLE 9.8 Overall accuracy, producer's accuracy, user's accuracy, and Kappa coefficient produced by the nearest neighbor classifier
with mean value of the original bands, brightness band, and maximum difference band.
Referece Data
Total
Producer's
User's
B
I
T
G
P
S
Sh
Pixels
Accuracy
Accuracy
B
33
2
1
4
1
0
0
41
91.67%
80.49%
I
2
35
9
1
0
8
6
61
85.37%
57.38%
T
0
1
12
5
0
0
0
18
50.00%
66.67%
G
0
0
0
11
0
0
1
12
47.83%
91.67%
P
0
0
0
0
14
0
2
116
93.33%
87.50%
S
1
1
2
2
0
25
2
33
76.76%
75.76%
Sh
0
2
0
0
0
0
0
17
60.71%
89.47%
Total Pixes
36
41
24
23
15
33
28
200
Overall Accuracy
73.50%
=
Kappa Coeficient
=
0.68%
sets of features. Differences in accuracies may partly be due
to the fact that different training samples were taken for both
sets of data. The same training samples should not be used for
different sets of band combinations as is normally done with
supervised per-pixel classifiers. This is because objects derived
from one set of features (different combination of bands) are
different from those generated by another set of features, even
though segmentation parameters (compactness, shape) and scale
parameter may be the same. The overall classification accuracy
could be increased by more effectively identifying swimming
pools using the rule set approach developed in this study and
overlaying the output pool map with the land-use and land-cover
map generated by nearest neighbor approach.
that represents reasonably the entire study area and all classes
of interest before performing the classification for the whole
data set. If one feels that a small subset may not represent most
of the study area, then several different small subsets could be
used to develop the segmentation and classification parameters
and procedures. However, selected decision rules will still need
to be tested for an entire dataset upon developing and testing
different decision rules. The threshold values and data ranges
used for different features may need to be modified for the entire
dataset, especially when implementing different decision rules
for purposes of separating overlapping classes.
A general procedure to perform the image classification phase
based on a rule set approach is:
1 Understand the overview of the classification specificity,
nature of classes and data, and potential signature confusion
among selected classes.
2 Select the image subset.
3 View image (e.g., waveband) features and develop a classifi-
cation strategy.
4 Transform the strategy into a rule set.
5 Classify the subset image.
6 Qualitatively evaluate the results.
Conclusion
The rule set classifier is generally more time consuming and
difficult to implement than identifying classes using a nearest
neighbor classifier. We recommend that different rule sets be
developed and that the effectiveness of potentially strong decision
rules be tested with a small subset. As demonstrated in this
chapter, it is a good idea to subset a small part of the data
Search WWH ::




Custom Search