Search

Neural networks represent a widely used alternative to deal with remotely sensed image data. The improvement of spatial and spectral resolution in latest-generation Earth observation instruments is expected to introduce extremely high computational requirements in neural network-based algorithms for classification of high-dimensional data sets such as hyperspectral images, with hundreds of spectral channels and very fine spatial resolution. A significant advantage of neural networks versus other types of processing algorithms for hyperspectral imaging is that they are inherently amenable for parallel implementation. As a result, they can benefit from advances in low-cost parallel computing architectures such as heterogeneous networks of computers, which have soon become a standard tool of choice for dealing with the massive amount of image data sets. In this chapter, several techniques for classification of hyperspectral imagery using neural networks are presented and discussed. Experimental results are provided from the viewpoint of both classification accuracy and parallel performance on a variety of parallel computing platforms, including two networks of workstations at University of Maryland and a massively parallel Beowulf cluster at NASA΄s Goddard Space Flight Center in Maryland. Two different application areas are addressed for demonstration: land-cover classification using hyperspectral data collected by NASA over the valley of Salinas, California, and urban city classification using data collected by the German Aerospace Agency (DLR) over the city of Pavia, Italy. [ABSTRACT FROM AUTHOR]

Hyperspectral imaging is a new technique in remote sensing that generates hundreds of images corresponding to different wavelength channels for the same area on the surface of the Earth. Most available techniques for hyperspectral image classification focus on analyzing the data without incorporating the spatial information; i.e. the data is treated not as an image but as an unordered listing of spectral measurements where the spatial coordinates can be shuffled arbitrarily without affecting the final analysis. Despite the growing interest in the development of techniques for interpretation and classification of such high-dimensional imagery, only a few efforts devoted to the design of parallel implementations exist in the open literature. In this paper, we describe AMEEPAR, a parallel morphological algorithm that integrates the spatial and spectral information. The algorithm has been specifically optimized in this work for execution on heterogeneous networks of workstations. The parallel properties and classification accuracy of the proposed approach are evaluated using four networks of workstations distributed among different locations, and a massively parallel Beowulf cluster at NASA's Goddard Space Flight Center. [ABSTRACT FROM AUTHOR]