Image Processing Reference
In-Depth Information
“exhibits potential to be implemented as a standard measure applicable to urban
ecology.” Special features of the study are:
•
Per pixel V-I-S plus water classification from 1:5,000 scale CIR photography
along transects outward from city center
TM per pixel classification of forest/woodland, sparse vegetation/grass, cleared
•
land/exposed soil, developed (impervious), and water
Constrained spectral unmixing of TM data
•
Several useful results came from the investigation: (a) distinctive zones from the inner
city to the urban fringe were quantified, as impervious surface gives way to vegetation,
(b) a “density of development index” was suggested based upon the impervious
fraction, and (c) a “stability index” was suggested based upon the soil fraction and the
notion that in a built environment, exposed soil indicates land conversion.
6.4.5
AVIRIS Data: Neural Network Analysis
Ridd et al. (
1997
) employed a neural network analysis on AVIRIS data of the
Pasadena area to detect and map eight V-I-S classes: green trees and shrubs (Vts),
green grass (Vgg), dry grass (Vgd), dark impervious (Id), gray impervious (Ig), bright
impervious (Ib), soil (S), and water (Wtr). The study area includes a portion of the
San Gabriel Mountains. The key elements of this investigation are:
•
Two hundred and twenty-four narrow spectral bands (10 nm) to improve sensi-
tivity to the nuances of environmental variety with 20 m spatial resolution
Several sub-classes of V-I-S composition felt to be significant features of the
•
area of study
A directed effort to identify the most diagnostic bands specific to the eight cover
•
types from AVIRIS data
An artificial neural network (ANN) classification technique to maximize separa-
•
tion of sub-classes
The first step was to select several optimal training sites in the field to represent
each cover type in order to generate a signature. Using GenIsis® software the
224-channel signature for each of the eight classes was displayed. Using the Similarity
Index Map module of GenIsis® the field site selections were refined and final signa-
tures set. Then, the eight most diagnostic AVIRIS channels were selected to separate
the classes. Table
6.6
shows the eight neural network channels and the associated
AVIRIS channel numbers. (Although AVIRIS channels are nominally 10 nm band
width, it was determined that they be doubled or even tripled for best discrimina-
tion.) The “color” range, band range, and band width are displayed. To reduce
massive processing time bands 5 and 6, and bands 7 and 8 were ratioed to result in
six input bands.
The neural network classification utilizes a “hidden layer” which allows multiple
combinations of linkages (synapses) between input vectors and output classes,
