993
a Signal and Images Processing Laboratory, Electronics and Computer Science Faculty, U. S. T. H. B., P. O. Box 32, El
- Alia, Bab - Ezzouar, 16111, Algiers, Algeria - s_chitroub@hotmail.com
b Al Madinah Remote Sensing Center, AlMadinah RD, Riaydh Alkhabra City, P.O.Box 242 Saudi Arabia -
ssulltan@hotmail. com
KEY WORDS: Vegetation analysis, PCA, Independent component analysis, Neural network, Change detection
ABSTRACT:
The seasonal analysis of vegetation can be considered as looking for fundamental redundant information and detecting, at the same
time, the natural changes of the vegetative cover undergone by the observed scene. From the statistical point of view, the redundant
information can be quantified by the correlation coefficients between the multi-temporal images while the natural changes can be
considered as the mutual information between the transition zones of the observed scene. For detecting and emerging the zones of
transition and preserving at the same time the zones of vegetation temporal evolution stability, it is interesting to create new images
in which the correlation between the images is vanished and the mutual information is minimized. To reach such purpose, we have
developed a new approach for seasonal vegetation analysis based on a new statistical multi-variate method called independent
component analysis (ICA).
1. INTRODUCTION
Multispectral image processing is a promising tool for the
analysis of vegetation in remote sensing imagery, particularly in
areas with low vegetation cover (Chen, 1998a ; Chen, 1998b ;
Gracia and Ustin, 2001 ; Shabanov et al., 2001; Kogan et al.,
2003; Frank and Mentz, 2003). Seasonal vegetation analysis in
the absence of land cover change is much more challenging
than land cover analysis. A variety of multispectral vegetation
indices have been developed in order to detect these changes
(Chen, 1998b ; Kogan et al., 2003; Frank and Mentz, 2003).
However, these indices are insufficient for seasonal vegetation
analysis, especially when the number of spectral bands is
important, in which make full use of the available spectral
images becomes impossible (Frank and Mentz, 2003). In
addition, the developed vegetation indices are limited in the
detection of low vegetation cover because of varying
background signals (Frank and Mentz, 2003).
The seasonal analysis of vegetation can be considered as
looking for fundamental redundant information, which exists
between the multi-temporal remote sensing data (acquired for
the same scene) and detecting and emerging, at the same time,
the natural changes of the vegetative cover undergone by the
observed scene. The redundant information characterizes the
stability in the vegetation evolution in the areas that are not
undergone to the natural changes across the time. The natural
changes, however, characterize the transitions across the time
between the states of the natural change zones of the scene.
From the statistical point of view, the redundant information
can be quantified by the correlation coefficients between the
multi-temporal images while the natural changes can be
considered as the mutual information between the transition
zones. For emerging the transition zones and preserving at the
same time the zones of vegetation temporal evolution stability,
it is interesting to create new images in which the correlation
between these images is vanished and the mutual information is
minimized. To reach such purpose, we have developed a new
approach of seasonal vegetation analysis based on a new
statistical multivariate method called independent component
analysis (ICA).
ICA is a useful extension of standard Principal Component
Analysis (PCA) (Chitroub, et al., 2001 ; Chitroub, et al., 2004 ;
Karhunen and Joutsensalo, 1994). As the name implies, ICA is
to find the transformation such that the resulting components
are as statistically independent from each other as possible. It
takes into account of higher order statistical properties and its
components are mutually independent with respect to these
higher order statistics, thus making ICA more truly independent
that PCA (Cardoso, 1999 ; Lee et al., 1999). The ICA model is
very suitable for neural network realization (Hyvarinen, 1999 ;
Lee et al., 2000). Most application of ICA so far has been on
Blind Signal Separation (BSS) of unknown source signals from
their linear mixture for which ICA obviously is useful. The use
of ICA for images has been much limited. We believe that ICA
can be useful in general in image and signal processing
(Chitroub, et al., 2004). In this paper, we will demonstrate some
potential advantages of ICA in remote sensing study. We are
concerned with the seasonal analysis of vegetation. The
remainder of this paper is organized as follows. The proposed
model is exposed in detail in section 2. Experiments performed
on the multi-temporal Landsat-TM images (they cover
AlQassim region in Saudi Arabia), are given and commented in
section 3. We conclude the paper in the last section.
2. ICA - BASED METHOD FOR SEASONAL
VEGETATION ANALYSIS
In this paper, we demonstrate the usefulness of ICA for
seasonal vegetation analysis. For that, a PCA-ICA neural