curve at band 34 (817 nm) is not normal, indicating a significant radiometric
problem. Inspection of the raw (i.e., non-calibrated) radiance image data showed
the same anomalous dip in band 34. A more extensive comparison of AISA
vegetation spectral signatures to vegetation signatures contained within spectral
libraries maintained at TEC revealed that bands 33 through 38 generally did not
provide quality reflectance data. This analysis was not quantitative, but was
based on extensive experience with analyzing vegetation signatures from more
mature hyperspectral imaging systems (e.g., AVIRIS, HYDICE) and from
ground-based vegetation signatures collected using hand-held spectroradiome-
ters. The decision was made to drop these bands from the images. Evaluation of
the remaining 32 bands suggested that many could likely be dropped without any
significant impact to image classification accuracy. Again, this was a qualitative
evaluation based on research team experience. Therefore, of the original 38
bands, a total of 14 were retained.
As stated above, signal-to-noise ratios for bands 1 and 38 severely dimin-
ished the quality of these bands. However, band 1 is the only channel close to the
center of the blue region of the spectrum (~450 nm) and was kept. Similarly,
band 38, while of relatively poor quality, was kept to at least provide a signature
end point well into the near infrared region. Figure 23 depicts a plot of the aver-
age DNs for BLINE7E and BLINE8W after this reduction in data dimensionality.
An additional benefit of reducing the dimensionality of the data was a sig-
nificant decrease in the file size. For example, the full 38 band flightline
BLINE7E is ~380 MB. After removal of 24 bands the file size is reduced to ~140
MB. By working with smaller files, subsequent image processing steps required
much less CPU time.
220.127.116.11 Cross-Track Illumination Correction. Test mosaics created using the
14-band, vegetation-only images still had pronounced seams where adjacent
flightlines overlapped. The mosaics showed that the south edge of each flightline
was consistently darker than the north edge of the overlapping flightline within
forested areas. An inverse shift in brighter to darker pixels was observed in the
water and marsh vegetation pixels, with brighter pixels at the lower edge of the
flightlines. This significant departure in the supposedly calibrated data was at-
tributed to cross-track illumination irregularities, which can be caused by any
number of factors, including:
Within sensor (i.e., system) radiometric distortions.
Bi-directional reflectance influences.