zetalemur
showcasing my project over the last years (again). The updated LC-Product works now pretty streamlined and it makes pretty interesting (and surprisingly good to compress) maps.
Stack is a collection of adjustable U-Nets that make predictions that are merged together to derive the final map.
Happy to answer questions.
Happy holidays and happy hacking.
It's also interesting to make maps directly from raw data (in this case that would be aerial imagery with mostly elevation data). E.g. if you find good training data you can easily make "land-cover" maps and although my example (of Austria) [1] does not have annotations (addresses, etc.), it's still nice to look at and has the benefit of being extremely compressible (compared to aerial raw imagery).
[0] https://wiki.openstreetmap.org/wiki/Map_features [1] https://turmfalke.httpd.app/test.html
AFAIK that is the common approach to solve this problem, but it still needs some degree of sensor accuracy?
Senior Software Engineer specialized in Django, Pipeline-Development, GIS and applied machine learning in the context of earth observation, sensor mapping in general and remote sensing in particular.
e-mail: [email protected]
Stack:
- Django
- Python (with eco-system: numpy, pytorch, etc.)
- Linux
- Godot
- GDAL
- OpenStreetMap
- Rust
- Bash-Voodoo
Available 20 to 100 hours/month (@110e VHB).Example work: https://turmfalke.httpd.app/test.html
Happy coding and nice to meet you!
Show HN: ~1m/Pixel EPSG 3857 Austrian LandCover Model
Hey HN,we developed a land cover model that works pretty well at least for Austria.Obviously it's not perfect and there is still lots of stuff to improve, but it's important to provide reproducible research artifacts in a world driven more and more by AI-based decision making.This is the first alpha release and the presentation needs some work, but the principal idea has been shown...
- 15000 Sentinel-2 scenes over land
- 10980^2 pixels per scene
- 50% lossless compression per scene
- 4 channels (RGB and infrared)
This results in 15000 * 10980^2 * 4 * 0.5 / 1024*4 ~ 3.3TB. This matches commercial offerings like in https://cloudless.eox.at/#data. Of course we have to add lower-scale imagery that is often used to provide XYZ tiling that would further increase the storage requirements.Thanks for the link to OpenTopography, it's an interesting project and elevation data is extremely valuable.
It might be feasible to store entire planet's raster tiles in 10 years in a (beefy) desktop machine.
So, good luck!
Sometimes you do not want to run an additional process/daemon like postgres. Your state is in essence now a file on some file system that you can atomically (full ACID) update using multi processing without the need for more complex machinery - you can get very far with this architecture.
I get your intention, but monetizing infrastructure and systems code is pretty hard.
If you need high accuracy, let's say for e.g. estimating eco system performance based on specific plant distribution, 75% is very low (especially if you want to feed it into another predictor) compared to a professional field biologist.
Why, though?
Is it because GHC's linear types are a superset of Rust's linear types? I guess that could rule out some features the compiler is able to prove (or not).
It's easily served by a simple server.
It makes sense, as (currently) it's pretty much impossible to tackle some important domains (e.g. image de/encoding) in Haskell while in Rust we already have libraries for that. Also: Idiomatic Rust is already extremely fast, while in Haskell it is non-trivial to implement high-performing code.
Prime example: The sieve of Eratosthenes (and I think we can agree that this is a prime example). In Rust you can implement a trivial solution that performs extremely well. In Haskell you will probably reach for the ST monad and I think it's controversial if that's idiomatic or not (and maybe non-trivial to grasp - as far as I know ST is implemented via compiler magic).
In my opinion the main culprit are modifier keys (e.g. by default the shell uses emacs bindings and you do a lot of ^A, ^E, ^R, etc.), but using jj mostly solves the issue as you just have to reach for Ctrl when you require SIGINT or copy/pasting in different environments that do not support vi keys. There are also people that remap it to Caps-Lock, but I just don't like to use my pinky that often.
Of course, everybody is different, but it makes sense to think about a setup that suits you and your health.
Part of the reason is probably that you need the proprietary drivers to use CUDA efficiently and as they are not in the (upstream) kernel you have to use some other tools, like unofficial repositories or the native installer (which doesn't always play well with the OS package manager, system upgrades, etc.). It's a real PITA.
However I learned to like it and use it often in CLI one-liners (mostly to cut out or reformat specific columns, probably the most common usage).