Hyperspectral vegetation identification utilizing polynomial fitting for dimensionality reduction

Abstract

Identification of vegetation species and type is important in many chemical, biological, radiological, nuclear, and explosive sensing applications. For instance, emergence of non-climax species in an area may be indicative of anthropogenic activity which can complement prompt signatures for underground nuclear explosion detection and localization. To explore signatures of underground nuclear explosions, we collected high spatial resolution (10 cm) hyperspectral data from an unmanned aerial system at a legacy underground nuclear explosion test site and its surrounds. These data consist of 274 visible and near-infrared wavebands over 4.3 km2 of high desert terrain along with high spatial resolution (2.5 cm) RGB context imagery. Previous work has shown that a vegetation spectral derivative can be more indicative of species than the measured value of each band. However, applying a spectral derivative amplifies any noise in the spectrum and reduces the benefit of the derivative analysis. Fitting the spectra with a polynomial can provide the slope information (derivative) without amplifying noise. In this work, we simultaneously capture slope and curvature information and reduce the dimensionality of remotely sensed hyperspectral imaging data. This is performed by employing a 2nd order polynomial fit across spectral bands of interest. We then compare the classification accuracy of a support vector machine classifier fit to the polynomial dimensionality reduction technique and the same support vector machine fit to the same number of components from principle component analysis.