search

allenai / satlas

Apache License 2.0
184 stars 19 forks source link

NAIP/Sentinel-2 joint training #22

Closed mizoru closed 9 months ago

mizoru commented 9 months ago

Were no models trained on the entirety of the dataset, that is both high-res and low-res images, and if so, why?

favyen2 commented 9 months ago

Not on the final dataset. We found that models inputting both high-res and low-res images did not benefit from the low-res images since the 10x difference in resolution means they don't really provide any additional information (even though they do have a few more bands).

We may explore training with both Sentinel-1 and Sentinel-2 images though.