Geoscience Reference
In-Depth Information
interpreter, an image processing specialist, a GIS specialist, and a statistician, conducted accuracy
assessments. A preliminary study was conducted, using data collected during a field trip to the
study area, to evaluate the effectiveness and accuracy of using aerial photographs to discriminate
grassland, desertscrub, and mesquite woodland classes. These classes were particularly difficult to
distinguish on the aerial photographs.
9.3.3.1
Image Collection, Preparation, and Site Selection
Landsat MSS data registration and other data integrity issues were reviewed for the 1973 and
1986 maps. These efforts included checking projection parameters and visual alignment using GIS
data layers (i.e., roads, streams, digital raster graphics, and digital elevation models). Random
sample points were generated using DOQQs acquired in 1992 (1:25,000 scale), and individual
sample points were located on the aerial photographs using the DOQQs for accurate placement. A
180-
180-m interpretation grid was generated and overlaid onto the LC maps.
Two mutually exclusive sets of sample points
¥
were generated for both 1973 and 1986 maps.
The second set of sample points served as a pool of substitute points when no aerial photographs
were available for a sample point in the first set. Whenever possible, pixels selected as sample sites
represented the center of a 3
3 pixel window representing a homogeneous cover type. For rare
classes (e.g., water), pixel sample points were chosen with at least six pixels in the window
belonging to the same class. A total of 813 reference samples were used to assess the 1973 (
¥
n
=
429) and 1986 (
= 384) maps. Multiple dates of aerial photographs were used in assessment: June
1971 and April 1972 (1973 map) and June 1983, June 1984, and September 1984 (1986 map).
n
9.3.3.2
Photograph Interpretation and Assessment
Photointerpreter training included using a subset of the generated sample points identified during
visits to the San Pedro watershed locations as interpretation keys. To avoid bias, photointerpreters
did not know what classifications had been assigned to sample points on the digital LC maps. To
locate the randomly chosen sample sites on the aerial photographs, the site locations were first
displayed on the DOQQ. Interpreters could then visually transfer the location of each site to the
appropriate photograph by matching identical spatial data such as roads, vegetation patterns, rock
outcrops, or other suitable features visible on the DOQQ and on the photograph. Each transferred
sample point was examined on stereoscopic photographs and identified using the definitions shown
in Table 9.1. LC categories for each sample point were recorded on a spreadsheet. A comment
column on the spreadsheet allowed the interpreter to enter any notes about the certainty or ambiguity
of the classification. The senior photointerpreter checked the accuracy of 10% of the sample point
locations and 15% of the spreadsheet entries to ensure completeness and consistency. All LC class
interpretations noted by a photointerpreter as “difficult” were classified by consensus opinion of
all the interpreters.
9.3.4
Digital Orthophoto Quadrangles
Approximately 60 panchromatic DOQQs acquired in 1992 for the U.S. portion of the study
area were available as reference data to evaluate the 1992 results. To obtain a precise geographic
matching between the DOQQs and the satellite-derived map, the 1992 source MSS image data
were geometrically registered to an orthorectified 1997 TM scene, and the resulting transformation
parameters were applied to the 1992 thematic map.
9.3.4.1
Interpreter Calibration
To effectively visualize conditions represented by the LC class descriptions (Table 9.1), Uni-
versity of Arizona and IMADES team members participated in a field visit to numerous sites in
