B13B-0567
Subspace method for multispectral, hyperspectral, and SAR image classification
Hasi Bagan, Yoshiki Yamagata
Center for Global Environmental Research, National Institute for Environmental Studies, Tsukuba, Japan

Various types of remotely sensed data (i.e., difference sensors, spatial resolutions, num of bands) collected over a span of years available today require advanced and innovative techniques to extract information and thematic maps useful for observing and characterizing both natural and human-induced changes over large areas. Land cover classification is one of the most important and typical applications of remote sensing. In the last years, new methods based on optimization and neural network algorithms have been proposed, and among them, subspace methods are very promising [1]-[3]. They are particularly useful for high dimensional and multisource data analysis, which is generally difficult to accomplish with classical statistical methods. The objective of subspace methods is to represent high-dimensional data in a low-dimensional subspaces. Classification then takes place on the chosen subspaces. In this paper, we carried out experiments with three types of remote sensing images, a Landsat ETM+ data, the AVIRIS hyperspectral data and CASI-3 hyperspectral data, and a recent lunched fully polarimetric Phased Array-type L-band Synthetic Aperture Radar (PALSAR) data. Experimental results show that subspace method is a valid and effective alternative to other pattern recognition approaches for the classification of various types of remote sensing data. The advantages of the subspace method are: (1) only 3 parameters are required to be set, and these can easily be determined by an automatic procedure; (2) The computational speed is faster than SVM and similar to the MLC; and (3) Could be a promising tool for future land cover classifications. References [1] H. Bagan, et al., “Classification of airborne hyperspectral data based on the average learning subspace method”, IEEE Geosci. Remote Sens. Lett., vol. 5, no. 3. 2008. [2] H. Bagan, Y. Yamagata. “Improved subspace classification method for multispectral remote sensing image classification”. PE & RS, vol. 76, No.11, 2010. [3] H. Bagan, T. Kinoshita, Y.Yamagata, “Combination of AVNIR-2, PALSAR, and Polarimetric Parameters for Land Cover Classification”, IEEE Trans. Geosci. Remote Sens. in press.


Fig. 1 Process flowchart of the subspace method for land cover classification.

Fig. 2 Top: plots of the accuracy rate vs. number of iterations. Bottom: corresponding classification maps.