Geoscience Reference
In-Depth Information
Table 9.5
Results of DOQQ-Based Accuracy Assessment of 1992 Land-Cover Classification:
Error Matrix and Producer's and User's Accuracy by Class
1992
Land-Cover
Classes
Reference (Digital Orthophoto Quads)
1
2
3
4
5
6
7
8
9
10
Grand Total
1
22
2
0
0
0
0
0
0
0
0
24
2
0
44
0
3
1
0
0
0
0
0
48
3
0
2
40
9
10
1
0
0
0
0
62
4
0
6
12
68
17
0
0
0
0
0
103
5
0
1
8
11
89
0
0
0
0
0
109
6
0
0
0
0
0
20
3
0
0
0
23
7
0
0
1
0
0
4
18
0
0
0
23
8
0
0
2
1
10
0
1
11
0
0
25
9
0
0
1
0
0
0
0
0
19
0
20
10
0
0
0
7
2
0
0
0
0
11
20
Grand Total
22
55
64
99
129
25
22
11
19
11
457
1992 Map
Total
DOQQ
Total
Number
Correct
Producer's
Accuracy (%)
User's
Accuracy (%)
Land-Cover Class
1. Forest
24
22
22
100
92
2. Woodland Oak
48
55
44
80
92
3. Woodland Mesquite
62
64
40
63
65
4. Grassland
103
99
68
69
66
5. Desertscrub
109
129
89
69
82
6. Riparian Forest
23
25
20
80
87
7. Agriculture
23
22
18
82
78
8. Urban
25
11
11
100
44
9. Water
20
19
19
100
95
10. Barren
20
11
11
100
55
Total
457
457
342
Note:
Overall accuracy = 75%; Tau = 0.72; Cohen's Kappa (Khat) = 0.70; standard error = 0.025.
and land-use practices. Classes mapped with lower than average accuracy included the small-area
agriculture, urban, water, and barren classes and the widespread mesquite woodland class. Factors
likely to have contributed to class confusions included: (1) LC changes between the dates of image
and reference data (especially for the 1973 and 1986 maps), (2) high spatial variability within
classes (including areas dominated by soil background reflectance), (3) variable interpretations of
class definitions by independent assessment teams, and (4) errors in reference data interpretation.
Geometric misregistration did not appear to be a factor in the results presented here.
The agriculture class had higher producer than user accuracies for all dates and was most
frequently confused with riparian, desertscrub, and mesquite woodland classes. The spatial distribu-
tion of agricultural areas in the watershed essentially outlined the riparian corridors, contributing to
mixed pixel spectral response and classification confusion. There may have been difficulty in distin-
guishing fallow and abandoned agricultural fields from adjacent desertscrub and mesquite woodland,
since the spectral response of these cover types was generally dominated by soil background.
The urban class included low-density settlement on both sides of the border. Low-density
development was difficult to distinguish from surrounding cover types even at the DOQQ scale,
suggesting the possibility of error in both maps and reference data. The accelerating pace of
development in the watershed, particularly in Arizona, may have contributed to cover changes
occurring between the dates of imagery and reference data.
The water class had the smallest area and was likely to have changed between the dates of
images and reference data, due to the ephemeral nature of most surface water in this semiarid
environment. For example, the 1973 NALC scene was acquired after a high-rainfall, El Niño-South-
ern Oscillation (ENSO) event during the winter of 1972-73 and portrayed wetter conditions than
