Geoscience Reference
In-Depth Information
Table 19.3 A Comparison of Accuracy Assessment Statistics Derived at Different Spatial Scales of
Analysis Using Per-Pixel and Whole-Area Assessment Comparisons of the Classified and
Planimetric Reference Data Sets Based on Simple Random Samples of the Data
Error
Matrix
Overall
Accuracy
(± 95% CI)
Error
Matrix
Overall
Accuracy
CV (%)
Portion of
Watershed
Sampled (with
Replacement)
Relative
Percentage
Correct
(± 95% CI)
Relative
Percentage
Correct
CV (%)
Pixels
Sampled
per Run
Area
Analyzed
(km
K
hat
(± 95%
CI)
K
CV
(%)
hat
2
)
Full
15,651
14.086
70.85 (-)
28.41 (-)
0.1853
(-)
Half
7,825
7.0425
70.61 (0.39)
1.65
28.31 (0.30)
1.31
0.1846
(0.0027)
1.85
Quarter
3,913
3.5217
71.18 (0.63)
2.74
28.72 (0.71)
3.08
0.1871
(0.0060)
3.98
Eighth
1,956
1.7604
72.33 (2.47)
11.16
29.12 (0.44)
1.87
0.1928
(0.0033)
2.11
Sixteenth
978
0.8802
71.67 (2.33)
10.26
27.74 (0.91)
4.12
0.1774
(0.0078)
5.48
1/25th
625
0.5625
67.20 (4.38)
16.67
27.79 (0.81)
3.65
0.1787
(0.0085)
5.93
1/40th
400
0.36
73.92 (3.06)
14.65
28.50 (2.13)
9.35
0.1899
(0.0197)
12.99
1/70th
225
0.2025
71.43 (5.74)
25.13
27.31 (2.43)
11.13
0.1780
(0.0244)
17.15
Note:
The relative percentage correct column lists the whole-area accuracy and the error matrix overall accuracy
and the Khat columns list pixel-based accuracy estimate values. CI = confidence interval; CV = coefficient
of variation.
for the entire Dead Run watershed was approximately 71% accurate. Also, the whole-area estimates
were robust with respect to the size of the sample subset, although the variability of the estimate
increased with smaller sample sizes. The per-pixel assessments of accuracy for the same unit-area
data sets were approximately 28% (Kappa 0.19) for the error matrix overall accuracy measurement.
For simple random sampling, the per-pixel assessments of accuracy showed less variability with
smaller sample sizes than the whole-area method, with the error matrix overall accuracy measure-
ment being particularly stable in this regard (Table 19.3). For pixel block sampling, measured
accuracy declined with smaller block sizes when considering both the whole area and per-pixel
methods of accuracy assessment (Table 19.4).
Table 19.4 A Comparison of Accuracy Assessment Statistics Derived at Different Spatial Scales of
Analysis Using Per-Pixel and Whole-Area Assessment Comparisons of the Classified and
Planimetric Reference Data Sets Based on Pixel Blocks Sampled without Replacement
Block Size
Sampled
(without
Replacement)
Area
Analyzed
(km
Relative
Percentage
Correct
(± 95% CI)
Error Matrix
Overall
Accuracy
(± 95% CI)
Pixels
Sampled
per Run
Error
Matrix
CV
K
hat
) per
Block
2
(± 95%
CI)
K
hat
CV
25
¥
25 blocks
625
0.5625
63.68
(13.43)
26.75 (4.48)
20.91
0.1593
(0.0439)
34.47
15
¥
15 blocks
225
0.2025
54.67
(13.33)
23.78 (8.42)
44.24
0.1214
(0.0751)
77.30
9
¥
9 blocks
81
0.0729
51.47
(17.21)
21.81 (8.04)
46.08
0.1018
(0.0730)
89.66
5
¥
5 blocks
25
0.0225
48.05
(26.33)
17.33 (7.75)
55.90
0.0455
(0.0576)
158.06
3
¥
3 blocks
9
0.0081
-17.81
(114.25)
18.52 (7.26)
48.99
0.0732
(0.0742)
126.71
Note:
The relative percentage correct column lists the whole-area accuracy and the error matrix overall
accuracy and the Khat columns list pixel-based accuracy estimate values. CI = confidence interval;
CV = coefficient of variation.
 
Search WWH ::




Custom Search