Michael Sumner

they wrote a geotransform with positive-y in *how much* data?

https://cloudnativegeo.org/blog/2026/02/the-technical-debt-of-earth-embedding-products

sheesh, an easy thing to have checked lol ... and now, rather than fix the sources they've forced a weird workaround in multiple libs that will confuse more processes and people in future ... #eyeroll

The Technical Debt of Earth Embedding Products

Every geospatial foundation model team solves the hard problem — training on petabytes of imagery. Nobody solves the easy one: letting other people use the output.

Apr 14 at 11:02pmWeb
Show threadMichael SumnerApr 14
no one in R probably would have noticed https://gist.github.com/mdsumner/6f592e6bb444aa974c2365ac435bb965 but sure, lecture us on technical debt after you made an enormous mistake
gist:6f592e6bb444aa974c2365ac435bb965

GitHub Gist: instantly share code, notes, and snippets.

Gist
Show threadMichael SumnerApr 14

when I use or create a new dataset I always check it carefully by reading it and making a couple of maps to verify alignment and sense

#thankyou,tedtalk

Show threadMichael SumnerApr 14

I guess this exciting new embeddings era with slightly funky transforms will flush out problems in these legacy downstream tools eventually

(ps, stop being so wedded to rasterio and other stuck-in-the-past python-only wrappers)

Show threadMichael SumnerApr 14
also that blog post is clearly drafted by an AI, not saying that's evil I'm definitely guilty of it - but you can see the tone after a long debug-slog, it's obvious