geo-deep-learning icon indicating copy to clipboard operation
geo-deep-learning copied to clipboard

Published 20 hours ago • verified publisher icon NRCan

Images_to_samples.py (hdf5s) --> data_to_tiles.py (tiles as tifs and geojsons)

Open remtav opened this issue 3 years ago • 3 comments

Features

  • CRS mismatch automatic handling
  • Resize image to new spatial resolution (ex.: 50cm to 25cm)
  • Don't rewrite tiles for an image that has all expected tiles
  • Crop geopackage to geojson (don't burn right away)
  • Input a csv directly (rush mode)
  • Input an image directly (for inference if pre-tiling is preferred?)
  • Parallelize tiling

Where to write tiles?

data_path / experiment_name / samples_folder_name ? ex.: geo_deep_learning/data/geobase_buildings/tiles_3bands_minannot1

How to deal with CRS mismatch?

  • to be discussed...

Onboarding solaris

  • How to integrate necessary solaris' utils?
  • 1st attempt

remtav avatar Dec 07 '21 20:12 remtav

Suggested logic for data_to_tiles.py

data_to_tiles_flow drawio

remtav avatar Jan 20 '22 19:01 remtav

Suggested class structure for use in data_to_tiles.py

classes_data_to_tile drawio

remtav avatar Jan 20 '22 19:01 remtav

Seeing comments from team members on Slack and Github, seems like this issue needs further discussion. I suggest starting from a pros/cons table, then discussing it in a meeting based on this table.

Tiling approach Pros Cons
HDF5 *
Plain .geotiff/.geojson

*This type of table that expects a list of items is much easier done with Google Doc than Markdown

@victorlazio109 , @mpelchat04 , @ymoisan , @CharlesAuthier, please add/review the following Google Doc "Pros and Cons" table. This will serve as a basis for discussing the relevance of this PR and overall transition away from HDF5s.

remtav avatar Jan 24 '22 15:01 remtav