A system for processing hyperspectral imagery: application to detecting forest species

Abstract

Although there are many approaches to hyperspectral data processing, they are typically based on an intuitive search of the appropriate spectral channels to solve the pattern recognition problem. To account for the accuracy of the computational procedures used, optimization techniques are needed to select the most useful spectral channels and to find contextual links for neighbouring pixels within a particular class of observed objects. We describe a system that merges both these types of mathematical formalism using a step-up method to extract the optimal channels from their entire set and to explain the contextual constraints on the images of high spatial resolution. The method is applied to forests of different species and age, which include areas illuminated by the Sun and shaded areas; these are the main classes recognized in this study. The proposed improvements in finding the specific information layers serve to enhance the computational efficiency of the techniques applied. These layers are formed by the sunlit forest canopy, sunlit background, and shaded background for a particular solar zenith angle during an aerial survey. The original system is created based on the relevant construction of the classifier employed, bearing in mind the signal to noise ratio of the hyperspectral device, its calibration, and the elaborated procedures of imagery processing. Results are shown of the related applications using the proposed system, which reveal the higher diversity in mapping forest classes due to the separation of the pixels in accordance with the indicated information layers. The accuracy of the pattern recognition for the processed scenes is shown to increase as the listed procedures are realized.