Geoscience Reference
In-Depth Information
Table 11.3
Typical Error Matrix
Classification of Pixels (ground truth)
Class
I
II
III
VI
V
VI
VII
VIII
IX
Total
Classification
I
14
3
7
2
5
7
38
of pixels
IV
4
8
3
2
1
3
21
(model)
V
2
16
1
19
VI
1
1
2
VII
1
3
3
7
VIII
2
2
IX
Total no. of
pixels
20
19
38
6
17
8
8
17
2
135
Overall classification accuracy 50.4%.
a
Example drawn from the classification of vegetation types from a satellite image. I = yertchuk, II =
gum/stringybark, III = silvertop ash, IV = blue-leaf stringybark, V = clearcut/road, VI = tea tree, VII =
gum/silvertop ash, VIII = black oak, IX = unclassified.
a
Ratio of the sum of correctly classified pixels in all classes to the total number of pixels tested.
taken in using systematic sampling and stratified systematic unaligned sam-
pling because these methods could overestimate population parameters. Con-
galton (1988) also stated that cluster sampling may be used, provided a small
number of pixels per cluster are selected (he suggested a maximum of 10 sam-
ple pixels per cluster). Stratified random sampling worked well and may be
used where small but important areas must be included in the sample. How-
ever, simple random sampling may be used in all situations.
The number of samples may be related to two factors in map accuracy
assessment: the number of samples that must be taken in order to reject a map
as being inaccurate and the number of samples required to determine the true
accuracy within some error bounds for a map.
Van Genderen et al. (1978) pointed out that we want to know, for a given
number of sample pixels, the probability of accepting an incorrect map. In
other words, when high mapping accuracy is obtained with a small sample
(e.g., 10 items), there is a chance that no pixels that are in error may be sam-
pled (i.e., a type II error
8
is committed). The corollary, as stated by Ginevan
(1979), is also important: The probability of rejecting a correct map (i.e., com-
mitting a type I error) must also be determined.
