Exploring Independent Component Analysis for Remote Sensing

This paper explores the utility of the Independent Component Analysis (ICA) for remote sensing applications and compares it directly to the Principal Component Transform (PCT). The ICA is a signal processing method that extracts independent signal sources from a composite signal. It has been applied to a wide area of applications ranging from electroencephalography, for understanding brain activity, to feature extraction, for edge detection and face recognition. Our supposition is that the technique has utility for multi-/hyper-spectral image analysis: Aspects of a scene may contain independent sources that arise from their spectral signatures. For example, mechanized vehicles under camouflage have an independent source signature compared to that of the background forestry. It is currently unclear how this independence can be separated using the ICA and what performance costs are involved. To focus our preliminary experiments, we base our study on a hyper-spectral image taken from the Hyper-spectral Digital Imagery Collection Experiment (HYDICE) sensor, an airborne imaging spectrometer. The image presents a foliated scene containing mechanized vehicles under camouflage. It provides 210 spectral bands between 400nm and 2.5microns. To compare the ICA and PCT we explore three metrics of information content: the signal-to-noise ratio, kurtosis, and entropy.