Rapid Feature Identification in Satellite Imagery with Reconfigurable Hardware

Abstract

Satellite image data consists of multispectral and hyperspectral data sets. More and more of these data sets are becoming available as a result of more commercial and military satellites being launched, and the instruments on the satellites having increased spatial resolution and an increase in the number of spectral channels. One of the challenges is how to process the large amount of data being generated and how to identify features of interest for further processing. There are many different approaches to accomplishing this. We use the k-means clustering algorithm. Clustering multispectral image data provides several important advantages. It provides a lossy compression of the data in the raw image, while at the same time organizing the data for further analysis. Compared to histogramming, clustering algorithms provide much more information about the content of the image. We are implementing k-means clustering on reconfigurable hardware. The k-means algorithm iterates over the image data until it converges, and involves a great deal of low-level pixel processing. For these reasons, it is very expensive to do in software. For the same reasons it is particularly well suited to a reconfigurable hardware implementation. However, implementing this and similar algorithms in hardware require different design tradeoffs than software implementations. In this talk, I will present the k-means clustering algorithm, its application to satellite image data, and its implementation in reconfigurable hardware. Our implementation has shown considerable speedup over the same algorithm implemented in software