High Spatial Resolution Digital Imagery
Classification of remotely sensed imagery, using either manual or automated
techniques, typically requires ground-truth information and accuracy assessment.
Ground-truth data quantifies vegetative parameters, such as species composition
and canopy densities, within discrete, geo-referenced sample plots throughout the
study area. These sample data statistically train the classification of the entire
data set. For image processing applications, ground-truth sample plots are pre-
cisely located within the imagery. Multivariate statistics (i.e., training statistics)
that define the unique spectral responses of the vegetative classes are extracted
from the imagery at these locations. Supervised classification algorithms, a sub-
set of a larger group of statistical techniques called discriminant analyses, use the
training data to assign each pixel to one of many discrete vegetation classes.
If training data are absent, more interpretive discriminant analysis techniques
are available to classify remotely sensed imagery into discrete land cover classes.
Unsupervised classification, or nonhierarchical clustering, is a standard image-
processing tool used to delineate spectrally unique feature classes, or clusters.
This technique is more subjective than the supervised technique, relying on im-
age analyst interpretive skill to accurately classify each spectrally unique cluster.
Poplar Island Unsupervised Classification
As the specific objective of the Poplar Island effort was to delineate only
three broad land cover classes without ground truth, a modified unsupervised
classification technique was employed. The modified approach included:
Segmenting each mosaic into two primary classes (vegetation vs. non-
vegetation) using the Normalized Difference Vegetation Index (NDVI).
Classifying each primary segment into eight land cover types using an
unsupervised clustering routine.
Combining the two cluster maps into a final classification.
The resulting class maps displayed 16 land cover types, ranging from deep
water (typically the darkest feature), to exposed tidal mud flats, to marsh vegeta-
tion of varying density and species composition, to evergreen forests, to sand
(typically the brightest feature).
No attempt was made to collapse the 16 classes into the three land cover
types required for the Poplar Island restoration effort. Because of the absence of
site-specific ground truth data, the 16 classes could be combined (i.e., recoded) in
various combinations to produce a final class map depicting water, intertidal