Georreferencing chips in a memory-efficient way

[lcoandrade]

Hi there!
I have a Kaggle notebook that is correctly georreferencing my predictions.

My predictions are georreferenced using Rasterio MemoryFile, but I need to keep my geo chips in memory to be able to use rasterior's merge function:

from rasterio.merge import merge

# Merge the chips using Rasterio's merge function
merged, merged_transform = merge(chips_list)

I would like to know if there is a way to use a generator to merge my georreferenced memory chips.

The idea is to be memory efficient.

Any ideias?

[lcoandrade]

Is there any official solution to make georreferenced predictions available now?

[adamjstewart]

Not yet, see #560 (comment) for my latest thoughts. There are quite a few issues holding this up:

• Add stitching utilities #30
• Add utility to compute bounding box of resulting prediction #35
• Store projection metadata in rtree index of GeoDataset #411
• Stitching: Access patch geo-transform in callback after predict #1407

This will likely become of more importance to my future research projects so I expect to personally work on this in the next year (no promises though).