The Köppen–Geiger climate classification made simpler (I hope so)

[TheSquareRootOf2]

I have thrown together a temperature generator, but instead of an expert systems/analytical approach I just built a neural net classifier for temperature bands that takes the elevation map as input and outputs January and July temperature band maps. I haven't put it out publicly on Github or anything yet because 1) it's pretty slow still (takes ~45 minutes to run on a 4000x2000 image) and 2) I wasn't sure under what conditions I'm allowed to distribute the training dataset I trained the classifier on (that's become more clear recently, but I'm somewhat novice at all the license details and so I've been reluctant to proceed).

That's an interesting approach! How did you manage to feed your neural network enough standardized temperature maps to make it effective?

There's also a software program called Clima-Sim, from Weather Graphics, which is aimed for simulating Earth but can apparently also be finagled to work for other arbitrary geographies and lets you edit a variety of relevant variables including axial tilt and so on. From what little I know about it it actually crunches a bunch of the equations to compute climatological predictions given those input items. That said, it's not free and I don't know all that much about it in detail, and it might be tricky to use in the sense of requiring a lot of climatological knowledge to get the best results. You can search it online if you're curious about it.

Yeah, I've heard of it and gave the demo version a try, but not being able to save your model is too much of a drawback for me. Essentially, this is a model like this -although simplified- I'd like to build. I have no deep climatological knowledge, but I've got some understanding of fluid mechanics and thermodynamics so I guess it's a start.

I have done some work on a mathematical model for temperature here. I took a lot of inspiration and knowledge from this thread but I'm not using GIMP or PS to do the actual work, so there would have to be some tweaks. But I've found this gives a reasonable baseline for temperature and a smoother, more precise transition between zones. Exact numbers can also help if for whatever reason you're interested in the temperatures at other times of the year and yearly precipitation/temperature values. I hope this helps somewhat!

Wow, how come I did not find that blog earlier? The research paper linked in the first article is also particularly interesting, since it pretty much follows the approach I had in mind and goes much further in depth than I could on my own.

[AzureWings]

That's an interesting approach! How did you manage to feed your neural network enough standardized temperature maps to make it effective?

Well, you're correct in noting that unlike a lot of machine leaning problems there's a very finite amount of training data available! A lot of what I've done is to try to extract additional input parameters from the input Earth data (which is just the elevation + land-vs-ocean map). For example, before running the classifier I compute two different measurements of the shortest distance from each land pixel to ocean (one weighting distances in a manner incorporating elevation and one that's just raw distance) and feed those in as additional inputs to the model. This has a visible impact when you look at my output as isotherms do "bend" in the middle of large landmasses similarly to how the real ones do in many cases. My results have a ways to go yet (notably they miss things like current influences and high-latitude temperature inversions), but do make clear isotherms while being a bit more organic than direct application of hard cutoffs. They're worse than I'd like but better than I'd feared (with only the one training set I'd worried about overfitting or just not having enough data for learning to begin with, since I can't exactly get more).

For an example, the result of running my classifier on this elevation map:


Was the following:

Northern hemisphere's summer ("July"):


Northern hemisphere's winter ("January"):


I also tried a similar approach for precipitation, but the results were far less satisfactory with just the set of derived input parameters (latitude, elevation, distance from ocean, and distance from ocean weighting changes in elevation as increased distance) I had already. What I'd like to add to the input parameters (for both temperature and precipitation) is some sort of "density of ocean vector" element that represents both the "net" direction to and how much ocean is close to a point, but I've been having trouble figuring out how to compute it efficiently (since at base the problem is aggregating all the ocean pixels for every land pixel; that's pretty much a non-starter unless I can determine a way to reuse enough of the work to avoid being O(x^2 * y^2)).

I have done some work on a mathematical model for temperature here.

Thanks for linking your blog - that's some amazing work. I especially like the looks you take at demographics and trade - those are things I was interested in simulating too but I had no idea at all where to start.

[srm038]

Thanks for linking your blog - that's some amazing work. I especially like the looks you take at demographics and trade - those are things I was interested in simulating too but I had no idea at all where to start.

Thanks!

Hopefully it sparks the imaginations and labor of others.

[Coggleton]

For the transition maps you suggest for precipitation (https://www.cartographersguild.com/s...l=1#post287571), for parts 3 and 4 you mention "Equatorward west side" for both - what does the "equatorward west side" refer to for zone 4?

[CTA]

Planning to run through this process soon--how do you make a world topographic map? Specifically, how do you figure out which regions are supposed to be at which heights? I currently have a map of my world's plate tectonics, but will I need anything else?

[Michi il Disperso]

Goodevening everyone!
I followed this tutorial for my brand new world, i found it really entertaining, especially guessing temperature and precipitation zones!
Now i'm struggling with the script for GIMP; i managed to run it after some problem, but now i'm stuck to an error: unorderable types: str() > float ()
Anyone can unravel it for me?
I'm rather ignorant in software in general..

Thanks!

Michele

[Charerg]

Goodevening everyone!
I followed this tutorial for my brand new world, i found it really entertaining, especially guessing temperature and precipitation zones!
Now i'm struggling with the script for GIMP; i managed to run it after some problem, but now i'm stuck to an error: unorderable types: str() > float ()
Anyone can unravel it for me?
I'm rather ignorant in software in general..

Thanks!

Michele

The GIMP script is made for GIMP 2.8, it probably won't work properly with a newer version of GIMP. You could either download the old version of GIMP to run the script, or use AzureWings' script (which doesn't require any program to be run, though it does involve installing Python).

[Michi il Disperso]

The GIMP script is made for GIMP 2.8, it probably won't work properly with a newer version of GIMP. You could either download the old version of GIMP to run the script, or use AzureWings' script (which doesn't require any program to be run, though it does involve installing Python).

Well, i said that badly, i actually used the AzureWing's one with Python

[AzureWings]

To clarify, are you trying to run the script via GIMP or on its own via Python from a command-line? The latter method's correct; the description of it in the first post as a "GIMP script" is a bit misleading.

[Azélor]

It's not a script made to run on Gimp?