Investigate automatic endmember selection

[jpolchlo]

There are various methodologies for determining interesting/representative spectra from a stream of candidates. We should investigate the lay of the land here and try out some options. This issue will be resolved by providing demonstration notebooks or other library code to do these endmember selection tasks.

[jpolchlo]

Feasibility study with pysptools and NFINDR

I have a rough demo notebook that uses some of the pre-baked methods provided by pysptools. Specifically, this uses the fairly standard approach of applying NFINDR to find, within the spectral space, extremal points (spectra) which, by linear combination, can be merged to describe the largest possible range of the observed spectra. This algorithm, after a deterministic initialization (using ATGP) has a stochastic component to it, and so we may not have found the truly best representative spectra, but the results are instructive.

By way of example, we'll consider the problem of detecting trash piles in landfills, with the hope of one day identifying uncontrolled dumping sites.

Consider El Sobrante landfill:

This is a large landfill area with a comparably small active area of dumping, circled in red.

We have an AVIRIS extract of this landfill to work from:

Using all the pixels in this extract area as input to the NFINDR algorithm, we find the following 10 spectra:

There is some marked similarity in some of the pairs of spectra, but mostly they are distinct. Taking these as a source for unmixing, using fully constrained least squares (all non-negative abundances, abundances sum to one), we see the following feature layers:

I've labelled each feature with a guess about what it is. Note row 2 column 4 is the closest to capturing the dumping area, which I've indicated in red. Bear in mind the AVIRIS image has a smaller active patch than the first image in this post.