Science developed a) machine learning model for spectral unmixing b) an atmospheric modelling pipeline that determines top of atmosphere radiance

Spectral unmixing good results for lab data, but very bad for atmospherically propagated datasets. Needed automation of ToA pipeline to generate more data. This has been (almost?) completed and is ready to go. Convert more lab data to ToA and run ML algorithms, expet to see good results for atmosohericallt simulatde data i.e. build / train the unmixing model. How long will this take to complete? Need to ask members, many are busy this summer with competitive job offers.

IS: would like this to be finished in summer.

"Proposal from Nooh & Ege":

1. members in Science are very capable, and want to work on advanced research tasks (i.e., don't want to work on boring stuff b/c competing offers from jobs). We need to retain talent, so the challenge isfinding projects that can keep them interested.
2. Problems FINCH "cancelled": hard to motivated re-doing science for a payload with even worse performance than we already have - it's already super constrained. Crop residue is not interesting with a cubesat. AAFC said "industry doesn't care about crop residue, only scientists".

Regarding new Mission Prospecting : optical payloads produce rather straightforward data, as opposed to e.g. SAR (easier to do data processing, perhaps instrument design is simpler). But, difficult to find an optical instrument that can produce a unique scientific value proposition because there are so many missions before us.

==> Side Quests: utilize expertise in ML for unmixing, super resolution for research paper work. Something like Aerospace Policy. Want to form a research team. Benefits from research could trickle into team. Can find a supervisor.

My response:

• I am not against this, research team is good. (as long as you Science & Data Processing) can keep your members entertained with problems to solve. However,
• I, however, REQUIRE all things for FINCH project to be finished.
  • unmixing model that works for top of atmosphere radiance
  • data handling pipeline and GUI, so that non-experienced user can do it.

====

Data Processing: can guarantee that what UTAT SS PAY requires gets done. Would like some independence for research team so they can find a professor supervisor.

• GOAL: BUILD LASTING ARTIFACTS. CONVERT PEOPLE-HOURS INTO DURABLE ENGINEERING PRODUCTS.
• we need a "GUI" that will work for anyone outside DP to be able to use.

Science : cannot guarantee that what UTAT SS PAY requires will get done.

• spectral unmixing / classification problems ARE TOO DIFFICULT TO GENERALIZE

If we are changing the instrument (changing focal length, spectral range, etc). then....

• option 1: we can find another (very simple) scientific end application, for example NDVI, algal blooms, etc. don't need a Science team to do this, could just get a single member & python function (valid).
• "not sure that your members would want to work this"?!
• it would be comical to downgrade from research to 3rd-year GIS student homework.
• sure, this is possible, but it's hard to motivate members to do this boring stuff.

Q: do you acknowledge that the goals of UTAT SS PAY have shifted from a research & scientific novelty goals to a student learning project, in order to ? Do your members recognize this

strategic shift that to get CUBICS finding we will pivot to a "student learning easy payload" instead of a "fancy custom scientifically novel payload", being understood that you can work on research (and encouraged to do so) SO LONG AS the minimum requirements for the core of the mission get done.

Question (for Ege):

1. what will you have your members working on, if you have your choice, from now to June 2027?

   • want to continue with research.
   • "foundational models": spectral data exists in a multidimensional latent space, can inform certain metrics that would help with unmixing (?).
   • continue researching different methods for spectral unmixing, unmixing models / unmixing architectures.
   • building a "universal" atmospheric radiative transfer model. Current options (DART, MODTRAN, etc.) are difficult to set up. Can we make it easier?
   • 50% deep learning

   Generalizable model is possible once you have a foundational model, interested in researching this next. keep working on research related to Earth Observation