Google Alphaearth – AI that maps the entire planet with patient accuracy | Korben
So Google has just released something completely crazy that will strongly change the way we observe our planet. Alphaearth FoundationsThis is the name of their AI, which plays virtual satellites and can create super detailed cards from any place on Earth, at any time. And when I say great in detail, I speak of an accuracy of 10 meters to 10 meters. You may see your garden from space!
The principle is quite brilliant, because instead of having a physical satellite taking photos, Alfaearth Mouline billions (!!) already existing images from Sentinel-2, Landat, radar data that passes clouds, 3D field models, climatic information … shortly, everything that pulls as geospatial data since 2017.
Christopher Brown by Google Deepmmmmo introduced this by mapping the world ”in any place and at any time“And it’s not Flan because the system can even create cards for data where there are no direct satellite images. Between observation or extrapolates when he lacks data, a little like you had a Google Earth machine.
Scientists tested the equator system where the persistent clouds usually prevent agricultural observations and Alphaearth managed to map agricultural land at various stages of development without waiting for the sky to appear.
The performance of this AI is also the most impressive. It reduces errors by 23.9% compared to other existing approaches, while 16 times less storage space. To give you an idea, Nicholas Murray from James Cook University in Australia explains that his team is usually going through ”tens for hundreds of days“To process satellite data before you can start creating cards. With Alphaearth it’s almost instant.
Google turned 1.4 billion satellite tracks per year in the Earth Engine, making it one of the largest data files of the genre ever created. And all of this is that it is free for academic and scientific research. Scientists from around the world can therefore use it to study deforestation, planning projects of clean energy, monitoring urban expansion or monitoring climate change.
In addition, it has already used more than 50 organizations around the world. For example, the UN uses it for food security, Mapbiomas in Brazil monitors the deforestation of the amazons and global ecosystems atlas cartography ecosystems ever listed, such as the desert hyper-arid or coastal areas.
If you are developers or researchers and want to play with it, Google integrated it into Earth Engine with Phew features. You can conduct research according to similarity (like ‘Find me all places that look like this cornfield“), Automatically detect changes between two data, or create accurate cards with much less training data than before.
The research work has just been published on the Arxiv (not yet reviewed, but hey, it is Google Deepmind anyway) and explains that this is the first approach that supports the continuous time to observe the country. In principle, you can apply for a card on any date, not just those where there was a satellite.
When I see it, I think that the future already exists & a billion; There is no need to wait for the satellite to go to the right place at the right time, no more problem with clouds that hide everything. There are only AI that Hona’s data and that you do exactly what you need.
AH and Google announced that this thing is part of their new “Google Earth AI” initiative, which combines all their geo -space models. Apparently they also want to combine Alphaearth with Gemini (their LLM) in the near future. We might like to discuss with AI that would know the condition of the whole planet at any time … it’s crazy!
For those who want to dig an item, everything is available on Google Earth Engine and Google Deepmind blog. And if you are a researcher, go for it, it’s free and it can really change the game for your projects!
Source