Image prosessing and supervised classification
The image was processed and analyzed using ERDAS Imagine software.
Prior to image classification, all ground cover features not associated
with undeveloped desert land were extracted from the scene. This
includes all impervious and landscaped surfaces associated with urban
areas and exposed soil related to industrial sites, clearings for new
urban development, and major disturbances (Ward et al. 2000). A
LANDISCOR color aerial photograph (2000) covering the study area was
used as an interpretive guide for ground features in order to aid in
extraction of urban features. A supervised classification procedure was
used in order to assign vegetation classes to the image pixels. This
method involves creation of spectral signatures for each candidate class
based on training sites in the field, which contain vegetation
indicative for each class. These class signatures are used as a
reference tool for the assigning of community types to pixels according
to maximum likelihood
Field data collection
A pilot study, which used field sampling locations used in the
TWINSPAN classification as training sites, yielded poor results, so
several corrective actions were instituted in order to increase
accuracy. Given the area of pixels (900 m2), the original field sample
plots were too small to use as training sites without significant risk
of contamination by other community types. Therefore, a separate effort
was made to collect training samples of each vegetation type over an
area encompassing multiple pixels, which were recorded on GPS and
designated during the training process. Since these samples formed the
reference source for all vegetation in the study area, each training
site was selected as an unambiguous representative of a given community
type. Multiple training sites, scattered across the landscape as much as
possible, were used for each reference signature. Whenever possible, a
minimum set of pixels equal to ten times the number of bands used, 70 in
this case, was utilized in order to create reference spectra, as
recommended by Congalton (1991).
Auxiliary data
The larger-scale thermal band 6, which has a resolution of 120 m
rather than 30 m inherent to the other bands, was dropped and replaced
with a Soil Adjusted Vegetation Index (SAVI) layer calculated from the
Landsat image. SAVI is a modification of the Normalized Difference
Vegetation Index (NDVI), which is commonly used to detect
photosynthetically active vegetation by virtue of relatively high
reflectance of near-infrared and low reflectance of visible red light.
SAVI includes a correction for soil reflectance, which is especially
useful given deserts’ high exposed soil coverage. Efforts were made to
make the soil substrate in each round of image classification as
constant as practical so that the vegetation would be the dissimilar
variable between pixels. Soil texture influences the scattering of
incident light so that larger particle soils provide more surfaces off
which light can reflect. GIS-based soil maps were obtained from the
Natural Resources Conservation Service, a division of the US Department
of Agriculture (Soil Survey Geographic [SSURGO] database 2002). These
maps were used to divide the total study area into sections based on
texture characteristics: sandy, loamy, clayey, and coarse particle
soils. An additional unlabeled class, roughly coinciding with the
shallow bedrock of mountainous areas, was divided further into sections
consisting of continuous patches for individualized treatment. The
intention of this step was to conglomerate sites with similar
reflectance features for a common classification effort. Each separate
patch was analyzed using unsupervised classification into eight classes.
In unsupervised classification, reference spectra are not determined by
the user; rather, the classifier groups pixels based on spectral
similarity inherent to the image itself with only the total number of
classes selected by the user. A GIS layer depicting local geology was
utilized in order to visually ascertain correspondence between
geological formations and the image classification. If there was a
correlation, the candidate area was split into separate parts; if there
was no correlation, the patch was retained whole. Next, separate patches
judged to be relatively self-similar were combined into a common view,
the classification was repeated, and similar areas were aggregated. This
process resulted in seven different study sections, each of which was
classified on its own with reference spectra derived from training sites
located within each section, if possible. If a hypothesized vegetation
type was not located during the training process, a signature from
another section was used, though this was necessary only a few
times.
Accuracy assessment
A random subset of these points was chosen to be surveyed in the
field. Registration of Landsat pixels is not perfect; image
rectification and restoration from raw data necessarily distorts actual
positioning of pixels to a slight degree. For this reason, points were
designated from clusters of similarly classed pixels. Coordinates were
chosen from each image section, which allowed for a separate accuracy
assessment for each section’s classification. Vegetation within a 20 m
radius was surveyed to determine the appropriate community type. Since
the pixel array represents a two-dimensional depiction of the landscape,
training site radius was lengthened on slopes to allow for a horizontal
distance of 20 m. Post-groundtruthing procedures were used in order to
maximize accuracy, including refinement of training areas, deletion of
classes found to be absent or rare in each study section, and
aggregation of classes lacking strong discrimination according to
groundtruthing results, followed by reclassification of the scene.
Accuracy assessment was reported using an error matrix (Congalton and
Green 1999). Overall accuracy is a holistic summary of how successful
predicted class membership agreed with field observations from the
groundtruthing effort, and is calculated as the sum of the diagonal
cells divided by the total survey sites used to assess that particular
classification. Producer’s accuracy demonstrates how well survey site
pixels of a particular vegetation type are classified, and equals the
number of correctly classified sites divided by the total number of
survey sites for that type, the column total (Lillesand and Kiefer
2000). User’s accuracy represents the probability that a classified
pixel indicates the correct vegetation type in the field, and equals the
number of correctly classified sites divided by the total number of
sites that actually belong to that class, the row total. Since accuracy
assessment was not feasible for the more remote or inaccessible
locations of the Sierra Estrella Mountains, the McDowell Mountains, and
the sandy soil of the Hassayampa River, these areas were not
conducted.