Brief Intro and Layout

This pipeline is absolutely principal in doing the actual linear inversion. Currently, we have all these scattered solutions to also scattered problems. We have to get these together, implement a smooth program where we can plug-in-play with different combinations of these solutions, this requires an architecture shown as below:

S(2).png

Current Issues and Room for Improvement

Issues With mrEM Determination

Right now, we do not have any other method than “closest to centroid” that might give the mrEM. More alternatives are desperately needed.

Issues With Abundance Estimation

SFA is incredibly bad. There are no current indications as to why it might be behaving this badly but for a lot of the spectra it outputs literal garbage.
RMSE gradient descent is suboptimal. It has really weird outputs where it outputs the abundances in these “bands”, it gives predictions only in these bands. This in my opinion (Ege) is an indicator that the gradient descent algorithm is not working just as intended such that the actual manifold that all of this gradient descent happens (where we have 2 axes representing 2 abundances and the z axis as the RMSE), has a bunch of local minima where the algorithm seems to get stuck at. To find out more about this, the team (Zachar) is working on a brute force method to explore how the manifold actually looks like.

Layout

Part 1: Data Separation

This part is for data separation. Here, we separate the data into training and validation where both training and validation contain endmembers and mixed spectra. There are options to choose how this separation happens and how much we separate. Not much more than this.