Geoscience Reference
In-Depth Information
Table 9.6
Results of Video-Based Accuracy Assessment of the 1997 Land-Cover
Classification: Error Matrix and User's and Producer's Accuracy by Class
1997
Land-Cover
Classes
Reference (Video Frame Data)
1
2
3
4
5
6
7
8
10
Grand Total
1
20
4
0
0
0
0
0
0
0
24
2
2
50
0
3
0
0
0
0
0
55
3
0
1
27
13
12
2
0
1
0
56
4
0
8
16
113
21
0
0
1
0
159
5
0
4
4
12
115
0
0
2
0
137
6
0
0
0
0
0
21
2
1
0
24
7
0
0
1
0
15
2
5
1
0
24
8
0
0
0
0
0
0
0
24
0
24
10
0
0
2
0
19
0
0
0
3
24
Grand Total
22
67
50
141
182
25
7
30
3
527
1997 Map
Total
Video
Total
Number
Correct
Producer's
Accuracy (%)
User's Accuracy
(%)
Land-Cover Class
1. Forest
24
22
20
91
83
2. Woodland Oak
55
67
50
75
91
3. Woodland Mesquite
56
50
27
54
48
4. Grassland
159
141
113
80
71
5. Desertscrub
137
182
115
63
84
6. Riparian
24
25
21
84
88
7. Agriculture
24
7
5
71
21
8. Urban
24
30
24
80
100
9. Water
N/A
N/A
N/A
N/A
N/A
10. Barren
24
3
3
100
13
Total
527
527
378
Note:
Overall accuracy = 72%; Tau = 0.68; Cohen's Kappa (Khat) = 0.65; standard error = 0.024.
reference aerial photography acquired in 1971 and 1972 (Easterling et al., 1996; NOAA, 2001).
The water class was not evaluated in 1986 and 1997 assessments due to insufficient representation
in reference data.
The barren class was mapped with poor accuracy overall, including 0% correct in 1986. This
class was most often confused with mesquite woodland, grassland, and desertscrub. These classes
generally have sparse vegetation cover, with many image pixels dominated by soil or rock spectral
responses, and were difficult to distinguish from truly barren areas at the MSS 60-m pixel size. A
total of 38% of samples interpreted as barren on reference aerial photography from 1971 and 1972
were mapped as water in 1973; this was probably due to the interannual variations in precipitation
mentioned above.
The mesquite woodland class may be interpreted as an indicator of landscape change in the
San Pedro Watershed (Kepner et al., 2000, 2002). Conversion of many grassland areas to shrub
dominance during the last 120 years is well documented for this region (Bahre, 1991, 1995; Wilson
et al., 2001), and these change detection results were of potential interest to many researchers.
However, both user and producer accuracies of all four dates were generally low for mesquite
woodland (30 and 80%, respectively, for 1973 and 40 to 65% for other years). Class confusions
included all but the forest class, with especially large errors in the grassland and desertscrub classes.
This result may substantially reflect both the spatially and temporally transitional nature of the
class and differences in interpretation among the groups performing image classification and
accuracy assessment. Additionally, it was likely that neither the spectral nor the spatial resolution
of MSS imagery was adequate to distinguish the mesquite woodland class in a heterogeneous
semiarid environment, where most pixels are mixtures of green and woody vegetation, standing
litter, and soils of varying brightness (Asner et al., 2000).
