Analysis of forest environments - classification as a metric of hyperspectral instrument performance

Abstract

In considering the design of an operational space-based hyperspectral imager, instrument characteristics such as signal-to-noise, ground resolution and spectral coverage are factors for both system capability and cost. To provide a basis for imager optimization, an exploratory study was performed to investigate the impact of instrument characteristics on forest species classification, as an example criterion. A study site with pure and mixed western hemlock and Douglas fir stands was imaged with Hyperion and AVIRIS. The data were analyzed using classification accuracy of a maximum a posteriori Bayesian classifier applied to selected maximum noise fraction (MNF) transformed features and a random subspace binary hierarchical classifier as a metric for instrument performance. Quantitative results for signal-to-noise, ground resolution, and spectral range suggest operational parameters for hyperspectral imaging systems and clearly indicate the need for advances in methodology for analysis of hyperspectral data.