Geoscience Reference
In-Depth Information
Table 9.4
Error Matrix Comparing Aerial Photo Interpretation and 1986 Land-Cover
Classification, with Producer's and User's Accuracy by Class
1986
Land-Cover
Classes
Reference (Aerial Photo Interpretation Class)
1
2
3
4
5
6
7
8
10
Grand Total
1
19
1
0
0
0
0
0
0
0
20
2
3
35
0
1
0
0
0
0
0
39
3
0
0
17
3
19
0
1
0
2
42
4
0
0
12
77
12
0
1
0
2
104
5
0
0
8
13
74
0
0
0
0
95
6
0
0
0
1
1
19
2
0
0
23
7
0
0
1
4
3
2
9
1
0
20
8
0
0
0
5
3
0
0
13
0
21
10
0
0
3
10
7
0
0
0
0
20
Grand Total
22
36
41
114
119
21
13
14
4
384
Producer's
Accuracy
(%)
User's
Accuracy
(%)
1986 Map
Total
Photointerpreter
Total
Number
Correct
Land-Cover Class
1. Forest
20
22
19
86
95
2. Woodland Oak
39
36
35
97
90
3. Woodland Mesquite
42
41
17
42
41
4. Grassland
104
114
77
68
74
5. Desertscrub
95
119
74
62
78
6. Riparian Forest
23
21
19
91
83
7. Agriculture
20
13
9
69
45
8. Urban
21
14
13
93
62
10. Barren
20
4
0
0
0
Total
384
384
263
Note:
Overall accuracy = 68%; Tau = 0.65; Cohen's Kappa (Khat) = 0.61; standard error = 0.029.
9.5 DISCUSSION
9.5.1
Map Accuracies
Statistics describing map accuracy were very similar among the four dates tested regardless of
differences in assessment methods and reference data. Overall map accuracies ranged from 67 to
75% and Tau values from 0.65 to 0.72. There were no statistically significant differences among
Khat values (0.61 to 0.70) for all possible date comparisons.
One aspect of sampling that differed among the assessments was the application of homogeneity
standards to the context of map sample points. Selection was made from the center of uniform 3
¥
3 pixel windows for the1973 and 1986 assessments, with an exception for rare cover classes requiring
only a majority of five or more pixels to match the center pixel. All sample points were selected
from uniform 3
3 windows in the 1997 assessment. In contrast, for the 1992 assessment, a map
class label was assigned as the majority of six or more pixels within a 3
¥
3 window centered on
the sample point. Although a positive bias may have been introduced by sampling only in homo-
geneous areas (Hammond and Verbyla, 1996), this effect was not apparent in results presented here.
¥
9.5.2
Class Confusion
For all dates evaluated the producer's and user's accuracies tended to be similar to the overall
classification accuracies and ranged between 61 and 100%. Generally low classification accuracies
were expected in a spatially heterogeneous setting such as the San Pedro watershed, where cover
types were distributed in a patchy fashion across the landscape due to climatic and edaphic effects
