High Spatial Resolution Digital Imagery
37
NIR - Red AISA(30) - AISA(22)
=
NDVI =
NIR + Red AISA(30) + AISA(22)
where:
AISA(30)
=
band 30 (782 nm)
AISA(22)
=
band 22 (671 nm).
The output floating point formatted image has a histogram range of 1 to +1.
Pixel values below 0 are generally associated with non-vegetated, or only
sparsely vegetated, surfaces. Such pixels include man-made features (e.g., roads,
roofs), bare earth, and water. As the NDVI values increase, so does the amount of
vegetative cover within each pixel to a theoretical maximum value of 1. Figure
22 shows the resulting NDVI image BLINE7E, as well as the original (i.e.,
floating point) histogram and the histogram after rescaling to an unsigned, 8-bit
radiometric resolution.
To create the land/water mask, the NDVI 8-bit file was segmented into two
new images, one with vegetated pixels and one with water pixels. The split was
determined by examining the distribution of the NDVI histogram. Ranges of
pixel values in the NDVI image associated with the major land cover classes,
including water, land, and transition areas, were determined through visual inter-
pretation. Then a precise breakpoint coinciding with the separation of land pixels
from water pixels was identified. In the case of BLINE7E, an acceptable break
was found at DN = 145 (Fig. 22). Finally, the NDVI image was split into two
distinct binary masks. The first mask represented water pixels (DNNDVI <= 145)
and the second represented land pixels (DNNDVI > 145). The land mask was ap-
plied to the full hyperspectral image such that only a subset of the original file
remained depicting only those pixels supporting vegetation. The accuracy of this
step was not quantified. However, a visual inspection was employed to determine
if water pixels remained in the output "land-only" file. The output image was
also inspected to determine if too many land pixels had been removed. If the
land-only image was not acceptable, the NDVI histogram was interpreted again
and a new breakpoint was selected. Across all flightlines, the breakpoints fell
between 135 and 165. Figure 21 shows the vegetated pixels segmented from
flightlines BLINE7E and BLINE8W. A plot of the mean DN's for all 38 bands is
also provided in Figure 23. All subsequent processing, including image classi-
fication, was performed on the vegetation-only images.
3.3.2.3 Reduction in Data Dimensionality. An interpretation of the mean per
band reflectance values, shown in Figure 23, suggests problems with data in the
near infrared region of the spectrum. Specifically, the pronounced dip in the
Previous Page
