Geoscience Reference
In-Depth Information
9.5.3
Future Research
Assessment of future LC classifications for the upper San Pedro area should incorporate some
measure of the reference data variability, perhaps also allowing a secondary class label (Zhu et al.,
2000; Yang et al., 2001). This may help to clarify the results for some cover classes. For example,
the low accuracies and class confusions associated with the mesquite woodland class may have
been due, in large part, to its gradational nature. If the interpreter would have been able to quantify
the confidence associated with reference point interpretations, there would not have been a need
to select sample points from homogeneous map areas, thus reducing the possibility of a positive
accuracy bias (Foody, 2002). Another useful tool for future San Pedro LC work is the map of all
sample points used in the accuracy assessment. Each point was attributed with geographic coordi-
nates and both map and reference data labels (Skirvin et al., 2000). These data could be applied
to generate a geographic representation of the continuous spatial distribution of LC errors (Kyri-
akidis and Dungan, 2001) to highlight especially difficult areas that should be field checked or
otherwise handled in the future.
9.6 CONCLUSIONS
The results discussed in this chapter indicate that historical aerial photography, DOQQ data,
and high-resolution airborne video data can be used successfully to perform classification accuracy
assessment on LC maps derived from historical satellite data. Archived aerial photographs may be
the only reference data available for retrospective analysis before 1992. However, their resolution
(1:40,000 scale for NAPP data) often makes this task difficult. Successful use of DOQQ data
requires precise geometric registration of the LC map to allow the overlay of orthorectified DOQQs.
The use of georeferenced high-resolution airborne videography as a proxy for actual ground
sampling in accuracy assessment provided the best method for current reference data development
in the San Pedro watershed. The advantages include: (1) cost-effective collection of a statistically
meaningful number of sample points, (2) effective control of coordinate locational error, and (3)
variable-scale videography that permits the identification of specific plant species or communities
of interest. Additionally, the videography provides a clear depiction of cultural features and land-
use activities. The main limitation of this method is that data are collected along predetermined
flight paths, thus constraining the sampling frame design.
9.7 SUMMARY
Because the rapidly growing archives of satellite remote sensing imagery now span decades,
there is increasing interest in the study of long-term regional LC change across multiple image
dates. However, temporally coincident ground-sampled data may not be available to perform an
independent accuracy assessment of the image-derived LC map products. This study explored the
feasibility of utilizing historical aerial photography, DOQQs, and high-resolution airborne color
video data to assess the accuracy of satellite-derived LC maps for the upper San Pedro River
watershed in southeastern Arizona and northeastern Sonora, Mexico. Satellite image data included
Landsat Multi-Spectral Scanner (MSS) and Landsat Thematic Mapper (TM) data acquired over an
approximately 25-year period. Four LC classifications were performed using three dates of MSS
imagery (1973, 1986, and 1992) and one TM image (1997). The TM imagery was aggraded from
30 to 60 m to match the coarser MSS pixel size.
A stratified random sampling design was incorporated with samples apportioned by LC area,
using a minimum sample size of
n
= 20 for rare classes. Results indicated similar map accuracies
