Data retiling

[meiertgrootes]

As team we need to be able to retile a Geotiff to a specified regular size subtile grid. This in order to feed images to the NN classifier.

• [x] preserve Geolocation information
• [ ] size should be conducive to CNN networks
• [x] Preferably reproject to EPSG:3031
• [x] Read one tile with a moving window, the windows have a constant overlapping with each other

[rogerkuou]

Reproject to a local system can be down via specifying the dst_crs as a Proj.4 string. Here is an example tutorial on doing this via pyproj. rioxarray also uses pyproj to do reprojection.

[rogerkuou]

We did several investigation based on this. The retiling method is in general based on how we want to do shuffling. If we do not aim at shuffling all windows for all tiles, we can load several (i.e. as many as memory allows) tiles in the memory, retile them to windows, and shuffle these windows. As shown in Francesco's Notebook Otherwise if shuffling for all windows is required, we can make a window list first and shuffle those. like here