New Google AI model maps world in 10-meter squares for machines to read Google has released a new AI model that maps the world in 10-meter squares for machines to read. The company's AlphaEarth Foundations model has been trained on vast amounts of Earth observation data from satellites and other sources to produce embeddings for computer programs, rather than visual imagery or some other form of …
Like always, the French have it solved a long time ago.
Its' called 'Grandeurs et unités - Systèmes d'unités pifométriques' ('Grossen und Einheiten - Nasimetrischeinheitensystem' in German, 'Quantities and units - Nosemetrical system of units' for the rest of the world).
So, if you're blessed enough to read some (pretty vernacular) French, here it is : https://asqualab.webmo.fr/documents/qualite/metrologie/19-unites_pifometriques.pdf
You're welcome.
Looking at Table S1 (pdf p.20) of the paper linked through "according" in TFA, it seems to me that they go from a 40-D spatial data vector (with maybe another 24-D of time-dependent Sentinel-2 & Landsat), and then run some magic ML through that to get a 64-D embedding on a sphere. The compression of source data features seems negligible (eg. 1:1), and a lot of the source data was already professionally preprocessed before its use here anyways (DEM, GEDI, ERA5, GRACE, NLCD, WikiPedia, GBIF), plus 10m is really the highest resolution of this dataset (available only for some optical bands of Sentinel-2 and the ice-monitoring SAR C-band of sentinel-1) ... most of the dataset is coarser (down to 100m).
Also, they appear to rely on some method of Dosovitskiy et al., 2020 for their analysis, which has never been peer-reviewed and published (it is a conference paper). So the whole approach and argumentation appears weak to me, and I don't really see anything that qualifies as results in there either (except for a feeling of "the whole earth at 10m, wow!" -- which this clearly isn't given the source data).
The description of what this analysis(?) software(?) database(?) is can probably do with a rewrite to enhance clarity imho. At present it looks to add zero value to the source datasets ... or maybe I misunderstood(?)
Ah, you forget the almost mystical ability of current AI to fill in the gaps. If we can "enhance" a blurry old TV programme to fully watchable modern standards Doom Watch style[1] then we can easily interpolate the missing 10m resolution data.
> most of the dataset is coarser (down to 100m)
After all, 10m from 100m, why, that is only a factor of ten, grab the polyfilla. What was that? Volume, not linear scaling? Ok, just a factor of 1000, still doable. Huh? 40 dimensions? Um, ah, - look, squirrel! (door slams)
> 10m is really the highest resolution of this dataset ... ice-monitoring SAR C-band of sentinel-1
"Gentlemen, Google Gemini assures us there are enough ice deposits below the Gobi Desert to defeat Global Warming. Acting on Gemini's information, we have started issuing shovels to the Bedouin."
[1] /s wrt the "enhancements", btw; cracking episode otherwise, especially wrt politicians.
