Robust successive projection for endmember selection

[jpolchlo]

Overview

For unmixing operations, one must possess an operable set of descriptive endmembers. Oftentimes, the best source for these endmembers is the scene itself, due to difficulties in acquiring sufficiently descriptive endmembers via other means (i.e., lab- or field-based spectrometers).

There are a variety of methods for determining which spectra to choose, but one class of algorithm (successive projection) relies on an iterative process of selecting a candidate spectrum and subsequently projecting the remaining spectra onto the subspace orthogonal to the chosen spectrum vector, repeating until the desired number of endmembers have been identified. This PR provides an implementation for one such algorithm—robust successive projection algorithm (RSPA)—that examines a group of potential endmembers at each stage and picks the one with the lowest scene-wide projection error, thus reducing the effects of outliers.

Closes #146
Closes #70

Demo

See the included notebook for an example.

Notes

This contribution adds a requirement for pysptools which provides other endmember extraction algorithms, including NFINDR, which is compared to RSPA in the demo notebook.

Checklist

• CI passes after rebase
• README.md updated if necessary to reflect the changes