Giters

wri / sentinel-tree-cover

Image segmentations of trees outside forest

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Bring small area S2 protocol up to large area protocol

JohnMBrandt opened this issue · comments

commented

and then re-run test metrics

commented

The changes include

  • Downloading S2 and S1 data in UTM epsg rather than WGS 84
  • Swapping the 5 band shadow calculation for a 3 band calculation
  • Switching cloud detection to the precomputed sentinel hub product
  • Downloading L2A 20m bands in 20 meter resolution, rather than 10
  • Downloading S1 in 5x10m resolution, rather than 5x5 meter
  • Switching whittaker smoother to use np.apply_along_axis
commented

Done, running tests.

commented

Done - the conclusions and follow ups are:

  • The new S1 Download properly clamps between 0-1, but old did not. Not sure the impact of this
  • The S2 tiles overlap within a one or two pixel margin
  • Should look more into S1 situations with >80 available imagery dates. I don't think the subsetting of imagery dates is proper (maybe mixup between ascending/descending paths?)