amount of vegetation. Examples of classes with the lowest NDVI values include:
open water, low density vegetation surrounded by water (e.g., Eleocharis), and
stands of dead Typha. The second segment, with NDVI values from the lower
middle portion of the histogram, included low to medium density vegetation
types, such as vegwater mixed and herbicide treated hyacinth. The third NDVI
segment comprised pixels with medium to high density vegetation, including
mixed vegetation, hyacinthgrass mix, and Phragmites. The last segment con-
tained the brightest (i.e., greatest green biomass) vegetation pixels. Pure water
hyacinth, dense areas of water lilies, and dense stands of smartweed are included
in the fourth NDVI segment. A total of 18 clusters were classified within each of
the four segments. The spectral signatures were plotted and, along with visual
interpretation of the original multispectral IKONOS scene, were used to assign a
descriptive class name to each cluster. Similar clusters were aggregated within
each segment. Then, the four segments were combined, with similar clusters
again aggregated. The final IKONOS derived vegetation map contained 18 the-
matic classes (Fig. 37).
220.127.116.11 Minimum Mapping Unit. A nine-pixel MMU threshold was applied to
the IKONOS class map to remove the thematic speckle and improve the utility of
the final product (Fig. 37).
Wetland Vegetation Mapping Over Blackwater Wildlife Refuge
The ENVI software routines for building training data sets to perform dis-
criminant analyses, including a variety of supervised classification algorithms,
extract reflectance value statistics using user defined polygons. These polygons,
called Regions of Interest (ROIs), can be of any size. The objective of extracting
DNs from a multi- or hyperspectral digital image is to define the n-dimensional
statistical space unique to individual, or groups, of separable surface elements.
Because training data sets must build multivariate statistics, including the mean,
the standard deviation, and the accompanying covariance matrix, for each indi-
vidual land cover class, ROIs generally encompass more than one single pixel.
Furthermore, multiple ROIs must be used, potentially throughout the entire im-
age space, to account for spatial variability within the scene.
5.3.1 Image Subsetting for Wetland Vegetation Mapping Over Blackwater
During several preliminary image classifications and signature analyses, pre-
formed on sets of adjacent flightlines as the larger mosaics were constructed, a
number of land cover classes were identified as potentially problematic. Specifi-
cally, the agricultural fields within and adjacent to the Refuge maintain spectral