5A.5
Hierarchical PCA Based Data Fusion
Abhishek Agarwal, Nortel Government Solutions, Lanham, MD; and H. EL-Askary, T. El-Ghazawi, M. Kafatos, and J. Le-Moigne
Abstract— Principal Component Analysis (PCA) has been widely used as a data reduction technique to overcome the curse of dimensionality [1]. Previously it has also been shown that PCA can be used as a data fusion tool [2] to fuse either spectral or spatial information. In this research we have shown that hierarchical PCA can be used to achieve better classification accuracy then direct PCA. In applying PCA, we combine the obtained information from the different angle views and frequency bands of MISR datasets for better and automatic dust cloud identification from other cloudy features that are often confused with dust. PCA is mainly performed in two general ways: directly and hierarchical. In the case of direct PCA, as shown before, we merge the 20 different images resulting from the different four spectral bands over the Nadir and the four forward angles. In the hierarchical case, we first merge the information from the 4 spectral camera bands separately in one step followed by merging the spatial information from the 5 cameras in the second step (or vise vera). The classification results over the original and the resulting data have been compared. Classifications by an expert Earth scientist were also obtained. The results have shown that fused data shows similar classification results for the Direct PCA methods where as using hierarchical PCA classification accuracy can be increased. This is attributed to the fact that applying PCA to the to one particular data domain (e.g. spectral followed by spatial data or vise versa) tend to better remove redundancies and enhance features in that given domain. Fig. 1 shows the comparison of classification experimental results obtained from original data as compared to direct PCA and hierarchical PCA. In addition, we will present that classifying through hierarchical data fusion as a first step, results in significant computational savings direct PCA.
Keywords- PCA; Data Fusion; MISR; Dust Storms.
References [1] S. Kaewpijit, J. Le Moigne, and T. El-Ghazawi, “Feature reduction of hyperspectral imagery using hybrid wavelet-principal component analysis,” Optical Engineering Vol 43 No 350, Feb 2004. [2] H. El-Askary, A. Agarwal, T. El-Ghazawi, M. Kafatos and J. Le Moigne, “Enhancing dust storm detection using PCA based data fusion,” Geoscience and Remote Sensing Symposium, IGARSS '05. Proceeding, vol 2, 25-29 July, PP:1424 – 1427, 2005.
Session 5A, Satellite IIPS and Applications
Wednesday, 17 January 2007, 8:30 AM-10:00 AM, 216AB
Previous paper Next paper