The International Archives oj the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
995
By this relationship, g(V) must be chosen so that its derivative
approximately forms a probability distribution function for the
sources to be recovered. The only remaining parameters to
adapt are the synaptic weights that can be found by maximizing
H(Z) with respect to B. The weight update rule will then be a
gradient descent in the direction of maximum joint entropy.
More computationally efficient approaches have been proposed
in (Lee et al., 2000), the reader can obtain more of the
mathematical details in (Chitroub et al., 2006). If we define the
term score function as:
^(v) = fep(v)/dv)/p(v) (7)
then an efficient weight update is:
AB cc (|-^(v).V 7 )b (8)
The form of q(y) plays a crucial role because it is function of
the transfer and therefore a function of the source estimate. For
the sub-Gaussian sources, the form of ^v) is such as:
(/Ay) = V - tanUy) (9)
where tanh(.) is the hyperbolic tangent. For the super-Gaussian
sources, (p(y) takes the form:
(p{y) = v + tanhiy) (10)
The switching between the sub-Gaussian and super-Gaussian
learning rule gives the following learning rule for the ICA -
part model (Lee et al., 1999):
AB oc (l - K.tanh(v).v T - v.v r )b (11)
^ is a A-dimensional diagonal matrix with elements
sigrifuiyi)) • ib(v,) is the kurtosis of the source estimate v,. The
switching parameter fo(v.) can be derived from the general
stability analysis of separating solutions (Cardoso, 1996;
Hyvarinen, 1999). 3
3. EXPERIMENTAL RESULTS
We present in this section our preliminary results. More
detailed study and more completed results are under
development. They will be subject of the future works for
publication. A real multi-temporal data provided by the
Landsat-TM are used to evaluate the proposed method. The
data were acquired over the AlQassim region in Saudi Arabia
(140x235pixels) during April and June 1994. The fifth bands of
the two sets of data are shown in Figure 4. The first three
extracted PC images are given in Figure 4. The first
components have the best image quality (contrast). Figure 5
shows the extracted IC images. These images are different to
the PC images. In the PC images the correlation is vanished and
consequently the redundant information is minimized. In these
images the zones of vegetation temporal evolution stability are
well mapped since they are characterized by the variance of the
PC image. This variance is maximized in the first PC images.
While the contrast between the input spectral images, which
characterize the differences between the spectral bands, is
mapped the last PC images that are much noised and
consequently it is not possible to overcome the information
about the natural changes undergone by the observed scene.
However, in IC images the mutual information between the PC
images are minimized and so the natural changes, which can be
considered as the mutual information between the transition
zones of the PC images, are emerged. In the IC images, the
zones of vegetation temporal evolution stability are also
preserved as they are in PC images. This can be quantified by
computing the image of the vegetation stability zones and the
image of the vegetation transition zones (natural changes) of the
scene both from the first and the second extracted IC images
(Figure 6).
4. CONCLUSION
In this paper, we have presented a new method for seasonal
vegetation analysis. The method is based on the emergent
technique, which is the independent component analysis (ICA).
The basic idea is to consider that the natural change undergone
by the observed scene can be detected and emerged if the
mutual information that exists between the images is minimized.
This task can be realised by the compound neural network
model PCA-ICA. We have presented here the main outlines of
the theoretical analysis of such model. More complicated
mathematical development of the proposed method cannot be
given here and it will be subject of the future works for
publication. Although the preliminary results, given in this
paper, needed to be evaluated by an objective and mathematical
criteria, they are, however, of interesting in the sense that the
extracted IC images are very informative about the surface state
of the observed scene, concerning the stability and transition
zones of the vegetation, compared to the extracted PC images.
REFEFRENCES
Cardoso, J. F. Laheldm B., 1996. Equivariant adaptive source
separation. IEEE Transactions on Signal Processing, 45(2), pp.
434-444.
Cardoso, J. F., 1999. High-order contrasts for independent
component analysis. Neural Computation, 11(1), pp. 157-192.
Chen, Z., Elvidge, C. D. and Groeneveld, D. P., 1998a.
Monitoring seasonal dynamics of arid land vegetation using
AVIRIS data. Remote Sensing of Environment, 65(2), pp. 255-
266.
Chen, Z., Elvidge, C. D. and Groeneveld, D. P., 1998b.
Vegetation change detection using high spectral resolution
vegetation indices. In Remote Sensing Change Detection:
Environmental Monitoring Applications and Methods, C. D.
Elvidge and E. R. Lunettat, Eds. Ann Arbor Press, Ch. 11, pp.
181-190.
Chitroub, S., Houacine, A. and Sansal, B., 2001. Neuronal
principal component analysis for an optimal representation of
multispectral images. Intelligent Data Analysis, International
Journal, 5(5), pp. 385-403.