Train on Geostationary Data #282
jacobbieker
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
One thing that could make Clay more useful would be to also train it on geostationary satellite imagery. That imagery has a much larger pixels (between 500m and 3km) that also differ in their spacing as you get further from the equator. On the other hand, the geostationary satellites have constant view of the whole disk visible to them, and with only ~5/6 geostationary satellites, you can cover nearly the whole world (GOES,MSG/MTG, and Himawari/G2KA). Training on this data could help with some other downstream tasks, such as renewable energy generation forecasting. While at Open Climate Fix, we were working on solar and wind forecasting, incorporating geostationary imagery to help "correct" NWP outputs of clouds and have more up-to-date cloud locations and forecasts of where they will move in the future. Most of the work was over the UK, and so only used a single satellite, EUMETSAT, but we wanted to bring it to many other parts of the world. Clay's approach to being able to take any input bands and varying amounts of inputs would be quite helpful there for a few reasons, specifically:
Geostationary imagery is also widely available, through Plantery Computer/AWS/GCP for GOES, Himawari, and G2KA. We also opened up 15 year archive of EUMETSAT data on Google Public Datasets too, in Zarr, to make it easier to work with for other ML people.
Beta Was this translation helpful? Give feedback.
All reactions