Code for generating benchmark

nasaharvest / crop-mask

End-to-end workflow for generating high resolution cropland maps

Apache License 2.0

94 stars 26 forks source link

Code for generating benchmark #380

Closed ivanzvonkov closed 3 months ago

ivanzvonkov commented 3 months ago

Adds a notebook to test different models against the representative test sets in the repository.

review-notebook-app[bot] commented 3 months ago

Check out this pull request on

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

ivanzvonkov commented 3 months ago

@gabrieltseng I am mainly looking for a second look at the Presto benchmarks at the bottom of the notebook, just to make sure everything is setup right.

gabrieltseng commented 3 months ago

The presto benchmarks look good to me :)

One note for the encodings - for the Taro mapping, results were better when we concatenated each channel group's embeddings, instead of taking a mean (link, but the code would look slightly different now because of the SRTM updates).

If we are confident that all the channel groups will be present then this might make sense to experiment with (but a drawback is that it makes the encodings quite a bit larger).

ivanzvonkov commented 3 months ago

Thanks @gabrieltseng, that suggestion definitely seems like the best next step!