Concurrent spatial-spectral band grouping: Providing a spatial context for spectral dimensionality reduction

This paper proposes a hyperspectral band grouping procedure that concurrently utilizes spatial and spectral information to identify an appropriate partitioning of the available electromagnetic spectrum. Spatial analysis is conducted per spectral band to provide a spatial context in which spectral information is utilized. The approach provides a means to exploit the natural variation of a groundcover´s spatial characteristics across the spectrum. This paper provides an overview of the proposed approach, details of various ways in which the approach can be implemented, and example hyperspectral analysis tasks where the method would be beneficial, as well as a detailed implementation of the approach for both a supervised and an unsupervised groundcover classification problem. The supervised and unsupervised groundcover classification methods are applied to airborne imagery, and experimental results are provided. Both methods´ results are highly promising. In particular the unsupervised groundcover classification method produces results on par with a highly supervised approach that has significant ground-truth requirements.