The V-I-S Model: Quantifying the Urban Environment - Remote Sensing of Urban and Suburban Areas - page 101

Table 6.6 Selection of AVIRIS channels resulting in eight ANN channels
NN channel AVIRIS channels “Color” range Band range (nm) Band width (nm)

1 12-13 Blue 508-518 20

2 29-31 Red 675-694 20

3 50-52 NIR 843-863 30

4 99-100 NIR 1,253-1,275 31

5 128-130 MIR 1,550-1,569 29

6 145-146 MIR 1,718-1,728 20

7 181-182 MIR 2,030-2,040 20

8 200-202 MIR 2,219-2,238 29

Fig. 6.14 Schematic neural classification from six spectral bands to an eight class map

until the optimal separability is reached. Figure 6.14 is a schematic diagram depicting
the role of the hidden layer leading to final classification. A pixel centered in a
3 × 3 × 6 matrix has multiple opportunities to be added to the hidden layer and
ultimately assigned to one of the eight V-I-S classes. Once assigned, the pixel takes
its place in the classification, in this case, class 3 - vegetation dry grass.
Figure 6.15 shows the overall accuracy in the training set at 86.8%, with Vgg
and Ib registering 100%, no errors of omission. Ig shows several errors of commis-
sion, confused with Id and soils. In the test set, overall accuracy was 89.95% with
Ig and soil registering 100% hits while Id missed about half of the actual Id pixels,
mistaking a large number as Ig. Because so little of the scene was water and in the
training set much of it remained unknown, it was omitted from the test set. It appears
from this study that AVIRIS 224-channel data coupled with ANN classification is