Effects of multispectral compression on machine exploitation

Conventionally, lossy compression techniques are evaluated in terms of standard performance metrics such as root mean square error and signal-to-noise ratio. It has become increasingly important in remote sensing applications to measure the impact of compression on the utility of multispectral imagery for machine-based exploitation. This paper describes a variety of metrics that measure compression effects on edge detection, clustering/classification, principal component analysis, and band ratioing. The compression algorithm employed is a transform coding technique using the Karhunen-Loeve transform (KLT) in the spectral domain, followed by a 2-dimensional discrete cosine transform (DCT) on individual eigen images. Results of applying this compression technique to a collection of ERIM's M7 multispectral imagery are presented. Performance measures in terms of both the conventional and machine-exploitation based metrics are also presented.<
>