Our research investigates methods to extract useful information from high dimensional hyperspectral data. Traditional cameras capture images in broad bands of red, green and blue (RGB) wavelengths. Benefiting from the advances of focal plane array technologies and gratings, these broadband systems now extend to hyperspectral imaging. In this scheme, individual bands (R, G, B) are captured across multiple channels, resulting in detailed information from the targets. Due to the nature of the information acquired (specific, narrow bandwidths detecting minor changes of reflectance), the datasets usually have intrinsic diagnostic characteristics with respect to the target. The identification of targets can be achieved using variations in spectral responses that can be attributed to situational, conditional, or illumination changes.

Most hyperspectral applications and assessments begin by addressing the dimensionality of the dataset. Spectral information is reduced using a preprocessing method to either identify features on the basis of interest or “project” them into a feature space based on the amount of valuable information. Even when attributing credibility to the narrow-bandwidth, contiguous bands for greater spectral information, the datasets also encounter the “curse of dimensionality.” Usually referred to as the Hughes Phenomenon, the statistical accuracy of class recognition is known to optimize for a subset of bands and subsequently decline due to inadequate training samples. The benefits of dimensionality reduction including the lower computational load are
widely accepted.

Our current research continues from our article (published in J. Appl. Remote Sensing) through a multi-University collaboration researching applications of the Airborne Prism EXperiment (APEX) data with Dr Yoann Altmann (Profile, Scholar) and Dr Andreas Hueni (Profile, Scholar).

Relevant Publications