Page 136 - Digital Analysis of Remotely Sensed Imagery

Image Pr ocessing Systems 107

Signature essential in supervised classification may be developed
from training samples or from laboratory spectral libraries. In addi-
tion, there are two novel classifiers in IDRISI Andes, the Fisher classi-
fier, based on linear discriminant analysis, and the backpropagation
neural network classifier. Apart from 13 such hard classifiers, there is
an extensive set of soft classifiers, totaling 14 for analyzing multispec-
tral data, such as those based on the Dempster-Shafer evidence the-
ory, fuzzy logic, and the linear mixture model. It is possible to combine
different classification procedures (e.g., Bayesian probability calcula-
tion with linear spectral unmixing) to form hybrid procedures to pro-
duce more reliable classifications. Also released in this version is the
largest suite of machine learning/neural network classifiers, such as
classification tree analysis, multilayer perceptron, self-organizing
feature map, and fuzzy ARTMap. Six special modules are designed to
analyze hyperspectral images.
IDRISI Andes is also able to carry out change detection based on
image differencing, image ratioing, time series Fourier analysis, spatial/
temporal correlation, and image profiling over time. Image differenc-
ing may be implemented via change vector analysis and regression-
based calibration. Derived using the temporal resonance module, the
temporal index indicates the degree of correlation between every pair
of pixels in multitemporal images. Special change analysis tools are
available for assessing change quickly. The most celebrated addition
to IDRISI Andes is the land change module for modeling ecological
sustainability, developed specifically for the International Center for
Biodiversity Conservation in the Andes. It is able to analyze land con-
version, predict and model change in the future via Markov chain
analysis or cellular automata, and assess the effect of the change on
biodiversity (Hermann, 2006). The modeled results may be validated
through a set of comparison tools with categorical map data. The
transition in land cover or potential of change can be explored from
both static and dynamic explanatory variables using either logistic
regression or multilayered perceptron neural network.

4.1.2 Display and Output
Raster images may be displayed in black and white or in true (24-bit)
color, or as transparent. Color display is possible with any bands des-
ignated as the red, green, and blue layers of a RGB composite
(Fig. 4.1). Each layer can be assigned a symbol file created with a spe-
cial symbol/palette development tool. Images may be displayed as
three-dimensional (3D) perspectives, contour plots, and analytical
hillshading. The fly-through module provides real-time interactive
animation over a digital elevation model (DEM). A 3D impression of
stereoscopic images may be obtained with the assistance of a pair of
anaglyphic glasses.
Moreover, images may be displayed on screen as a map composition,
into which nonimage layers such as hydrography, roads, and elevation