High Spatial Resolution Digital Imagery
91
ing valuable invasive species monitoring data by mapping the distribution of
hyacinth and lettuce over the whole of Lake Okeechobee.
9.4
Wetland Vegetation Mapping Over Blackwater Wildlife Refuge
The objective of this image processing research effort was to create an accu-
rate thematic map depicting detailed vegetation classes throughout the Blackwa-
ter Wildlife Refuge. A preliminary evaluation of the AISA hyperspectral imagery
suggested that association level classification would be feasible, considering the
high spatial resolution (3 m) and the excellent geo-registration of the 20 flightli-
nes needed to cover the Refuge. However, as the individual flightlines were as-
sembled to form a complete mosaic, serious radiometric problems were discov-
ered. Much of the research effort was spent trying to remove a significant cross-
track illumination anomaly and attempting to match, or balance, the histograms
of adjacent, overlapping flightlines to remove the visible seams. The final mosaic
was of acceptable quality for image classification, but retained some radiometric
distortions, including several significant seams and a number of holidays (i.e.,
holes in the mosaic).
The iterative spectral analyses employed to generate training signatures for
classification showed that stands of wetland plants, especially different reeds,
rushes, and vertical habit grasses, were spectrally inseparable. This observation is
not unexpected. Past research, where spectral signatures of vegetation were col-
lected using ground-based spectroradiometers, under natural illumination, con-
cluded that reflectance properties of unique vegetation types do not depend on
species, but rather on the plant phenology and the types of background materials.
Within the AISA imagery, a variety of factors influence the spectral characteris-
tics of the selected ROIs, including above-ground biomass, plant distribution
within the sensor field of view, crown and foliar structure, within and between-
plant shadowing, and the background soil properties. With such a wide range of
interacting physical parameters, many ROI signatures were not separable within
one standard deviation of their means. For example, deciduous tree crowns were
spectrally similar to dense grass cover or dense broad-leafed crops, and dead or
dying forests were spectrally confused with wetland grasses or reeds that support
necrotic stems and leaves. This inseparability produced a vegetation classifica-
tion with a relatively low overall accuracy.
The high spatial resolution acquired by the AISA sensor for this project
would suggest the potential for classifying the wetland vegetation classes with a
greater level of detail and accuracy. The marginal quality of the final map sug-
gests that smaller pixels do not necessarily result in a better classification. The
highly variable spectral response of adjacent pixels within the homogenous