There’s a new project that popped up recently that has published a global imagery of 2023 Sentinel 2 Imagery with a ‘zoom and enhance’ AI model applied to it to increase the resolution: SATLAS Super Resolution Map
Probably not good enough for tracing small buildings, but definitely good enough to check for the existence of new streets or large infrastructure changes.
It’s published under the ODC-BY license – I couldn’t find anything documented in relation to ODC-BY compatibility with OSM, but sounds like it’s maybe similar to CCBY4?
If nothing else, certainly helpful for comparing to other imagery and determining how up to date it is.
Edit: Though sadly, it appears they are using OSM w/o attribution as their base map for some of their other data sets.
I fail to see how using data that an AI hallucinated brings any benefit at all. Never ever use that for updating OSM! If the original data isn’t good enough to detect the “existence of new streets or large infrastructure changes” then the invented images aren’t either.
Please also tell this to the guys at Microsoft Buildings Footprints, especially those who click without checking what they are doing and end up adding buildings in the middle of rivers and lakes, or creating buildings in the position of a tree.
You’re right, sorry, my excitement about a recent global imagery set made me forget OSM’s cardinal rule: “dismiss anything with the word AI applied to it without a look”
This imagery set may not be it for the reasons you state, but does anyone know of an easily available imagery set created from recent Sentinel-2 data that is suitable for OSM editing? I’ve seen a couple of guides for downloading the data from Copernicus and creating tiles with QGIS – but it’s certainly a bit of process, if you are just trying to confirm if what Bing or ESRI is showing is still accurate.
Anyway, for now I’ll be out mapping with my sextant.
As you all know, most of these edits are done through Rapid and JOSM. That most mappers who do it have enough experience, knowledge and time in OSM to link the resources to mapping tools. Most of them are not newbies, when you write to them indicating that they have created a non-existent building they simply blame Microsoft Building Footprints, the JOSM Plugin or Rapid. But they will never admit that they have clicked on a lot of buildings that they have not verified.
When I find these non-existent buildings I simply delete them, I no longer waste time writing in the CS.
At first I was excited by the title of this thread, but unfortunately it must be admitted that they are in Beta phase, I extracted the URL of the images to test them and make comparisons.
The documentation says that the AI analyzes several images of the same area.
The first thing that stands out is the modification of Linear Geometries (roads and rivers). I thought it would be a substitute for mapping forest stands whose boundaries are not easily determinable when under cloud cover. But after several comparisons against Sentinel, Bing Aerial and ESRI images I think that at this stage it is not possible to use them. Maybe pick up again in a few months to see how it goes evolved the tool.
I share your scepticizm about AI but it’s well worth to have a look what this page is offering. I have checked a few places I know quite well and it’s amazing what they are able to squeeze out of the sat images. No fantasy entries detected so far but very actual display of new objects.
We do have residential construction areas all around and the images are good enough to crosscheck the details I have mapped OTG with my sextant … . And also to see if there are new buildings under construction which I have not yet mapped.
In another area extensive excavation is done to reshape a river/lake habitat. Very difficult to check OTG as part of it is prohibited to enter. The image created for this area looks very good to me, very close to what I have estimated so far. I’d say it’s good enough to map, at least until actual DOP20 images are available.
From what I have seen, I have to agree to
Would be great if they were going to update this every 2-3 month …
The Romaine-4 Reservoir north of Québec was flooded from 2021 to 2022 for the Romaine-4 Power plant. The AI model seems to have difficulty to take account of this. See satlas - Romaine-4 reservoir were you can compare AI Super Resolution and Tree Cover with OSM that was edited using Maxar, Esri and Sentinel-2021 Imageries. The AI model represents the two islands in the reservoir but with building style artefacts. A lot of artefactcs are also observed in the water surrounding the islands, what we dont see on the Sentinel-2021 imagery.
That’s quite possibly true - but if they don’t repeat the same mistake, that’s a result, even without a “mea culpa” reply.
A changeset discussion comment is useful to more than just the mapper whose changeset you’re commenting on. It’s useful to other mappers, as it suggests something that perhaps they ought to look for too. If they DO keep repeating repeating the same mistake, and a complaint comes in to the DWG, then it’s useful to us too.
I just checked this for some areas I work in, specifically railway yards in Mexico. The “enhanced” images are very inaccurate. Railway yards look like patches of concrete with some ridges that do not correspond to the real tracks. There is no concrete in these yards.
There is no “increase in resolution” here, only the illusion of increase of resolution.