198 CHAPTER 5. NEURAL NETWORKS FOR CONTROL
synthesis and chlorophyll production. Unhealthy plants, however, reflect
most of the radiation from the visible spectrum and less from the near-
infrared spectral band. The followingfigure illustrates the overall concept
of an ecosystem classifier.
To obtain an estimate of plant health, we compute a value called ìnor-
malized difference vegetation index (NDVI)î. It is a simple formula using
two satellite channel bands. For example, if one band is in the visible re-
gion (VIS) say Band 1 ( 0. 58 μmó 0. 68 μm), and one is in the near-infrared
(NIR) say Band 2 ( 0. 725 μmó 1. 1 μm), then
NDVI=(NIR−VIS)/(NIR+VIS)
The possible range of values is between− 1 and 1 , but the typical range
is between about− 0. 1 (NIR<VIS for not very green area) to 0. 6 (NIR
>VIS for very green area). Healthy vegetation reflects very well in the
near-infrared part of the spectrum. Green leaves have a reflectance of20%
or less in the 0. 5 to 0. 7 micron range (green to red) and about60%in the
0. 7 to 1. 3 micron range (near-infrared). Changes in the NDVI values help
in monitoring the health of the vegetation over time. NDVI therefore is
a very important index to estimate both the short-term as well as the
long-term health of agricultural ecosystems.
In this project you are required to train a neural network to classify veg-
etation. There are several Internet resources that can be useful.
ïwwwsgi.ursus.maine.edu/gisweb/spatdb/acsm/ac94014.html
A technical paper on this website entitled ìClassification of Mul-
tispectral, Multitemporal, Multisource Spatial Data Using Artificial
Neural Networksî provides considerable insight into the possible clas-
sifications of vegetation.
ïwww.geovista.psu.edu/sites/geocomp99/Gc99/072/gc_072.htm
This website has an excellent technical presentation by David Land-
grebe entitled ìOn information extraction principles for hyperspec-
tral dataî and provides a link to download free software.
A software application program called MultiSpec is available at no
cost from dynamo.ecn.purdue.edu/~biehl/MultiSpec/. It contains all
the necessary algorithms. Complete documentation of the program is
also available. The site also provides sample data sets. The program
can be used to extract features required for training a neural network.
