Fusion and Classification of Multisource Remote Sensing by ltx81750


									  Fusion and Classification of Multisource Remote Sensing Data
             Based on Tree Structured Filter Banks
                               J. R. Sveinsson and J. A. Benediktsson
                         Department of Electrical and Computer Engineering
                                        University of Iceland
                             Hjardarhagi 2-6, Reykjavik, IS-107, Iceland

   The application of neural network classifiers have given a good performance in fusion of remote
sensing and geographic data from multiple data sources. By using neural network approaches, the
achieved classification accuracies have improved significantly when compared to accuracies obtained
by the use of single data sources. However, there are several things that need to be considered when
neural networks are applied. For example, the representation of input data is important and can
significantly affect the classification performance. Feature extraction can thus be used to transform
the input data and find the best input representation for neural networks.
   In this paper two feature extraction methods are considered for neural networks used in classifi-
cation of multisource geographic and remote sensing data. The first method is based on the wavelet
packet transformation which transforms a signal from the time domain to the scale-frequency do-
main and is computed at several levels (tree structured) with different time/scale-frequency reso-
lutions. This transformation is a fast algorithm which splits the signal into subspaces or packets.
These packets represent different properties of the signal and hence some of them increase its fea-
ture extraction capability and others do not. Dimension reduction (or feature extraction) can be
achieved by throwing out packets which do not enhance the feature extraction capability. Hence,
this feature extraction method should be considered attractive for neural networks. The second
method is based on tree structured multirate filter banks. Tree structured filter banks are compu-
tationally efficient and can specifically be tailored for time-frequency analysis of multisource remote
sensing and geographic data. They can easily be adapted to different sampling rates and should
be considered attractive for feature extraction of multisource remote sensing data.
   For both of the considered feature extraction and image fusion methods a basis selection method
is examined, and its properties are utilized in the feature reduction schemes. The basis selection
method determines the shape of the transformation tree, evaluated from some specified criterion.
The criterion involves finding the maximum of the cross correlation between the input signal at level
L and the output signal at level L + 1 of the transformation. This criterion is a good measurement
of the similarity of the transformed signal at different transformation level. The criterion is such
that if the measure of the input signal at level L is greater the measure of the output signal at
level L + 1 the transformation is continued otherwise it is stopped at level L. Several criteria like
Lp-norm, maximum entropy, and minimum description length are considered in experiments where
the proposed feature extraction and image fusion methods for neural networks were used to classify
a multisource data set including multispectral remote sensing data, SAR data, and topographic
data. The performance of the proposed feature extraction methods is compared to results obtained
when conventional methods like principal component and discriminate analysis were used as feature
extraction methods for neural networks.

                             Progress in Electromagnetics Research Symposium, July 5–14, 2000, Cambridge, MA, USA

To top