High Spatial Resolution Digital Imagery
71
6
ACCURACY ASSESSMENT
Quantitative thematic accuracy assessment is critical in determining the
quality of vegetation classification. Typically, independent field data are used to
calculate thematic accuracy.
6.1
Accuracy Assessment Lake Okeechobee
The field data sample size was inadequate to provide for a thorough assess-
ment of the thematic accuracy of the either the full resolution or MMU filtered
vegetation class maps. However, a pseudo-accuracy calculation was completed
using the actual training site locations. This step is typically a preliminary proce-
dure in assessing the quality of the thematic map. In this case, the accuracy of the
24 class, MMU filtered map was ~ 80%. The aggregated 18 class, MMU filtered
map was only ~70% correct. These results suggest that a number of the sample
sites were incorrectly classified.
For example, the areas of mixed hyacinth are classified as Phragmites and/or
Typha/Mix. Also, much of the Phragmites classified throughout Moonshine Bay
(the southern most mosaic) is actually mixed vegetation. A more thorough visual
interpretation of the thematic product, combined with additional analyses of the
inseparability among the training signatures, suggests that the final CAMIS-
derived vegetation class map is no greater than 60% accurate.
6.2
Accuracy Assessment for Wetland Vegetation Mapping Over
Blackwater Wildlife Refuge
Owing to budgetary and time constraints associated with this effort, no an-
cillary field points were available. Instead, thematic accuracy was determined
using the training data field sites. Training sites are commonly used in this man-
ner only as a preliminary step in assessing classification accuracy.
The training sites used for classification were initially associated with very
detailed vegetative parameters, including species composition, species density
and background material. However, because of the relative lack of geometric
accuracy in locating the 1-m2 quadrats on the imagery combined with the mar-
ginal spectral separability of all of the class spectra, the final southern mosaic
class map was recoded to only nine vegetation classes. The geo-registered train-
ing sites for each of the mosaics were similarly aggregated to the same classes.
Image processing software generally offers accuracy assessment routines. In
this case, ERDAS Imagine was used to overlay a single point, generally the cen-
ter pixel of the training area, onto the renamed class maps. The thematic class
Previous Page
