HN Reader

Regarding the discussion of ger/yurt districts in cities, it's also important not to underestimate the cultural significance of the nomadic lifestyle and yurt culture.

Changing climate (desertification) and economic conditions have meant that a lot of people have given up their nomadic lifestyle and moved to cities or their outskirts (mostly Ulaanbaatar). They often are reluctant to do so, it's a big step, and they often hope it is a temporary one.

They set up their yurts not only because of housing shortages, but many are also hesitant to move into apartments or other permanent structures as it's seen as the last step in giving up this nomadic lifestyle. Often they are setting up their yurts next to permanent structures, either because they are living in the 'yard' of relatives or to expand their residences and stay connected to their culture.

You can see examples of this in the first images.

The gers are standardized. There is a big daily market in Ulaanbaatar where you can get all spare parts and complete gers. In 2017, the price for one ger was something like $1000.

For that money, you get a well-isolated easily movable tiny house in a country where you are allowed to settle everywhere (but if you have 2000 sheep with you, you should better discuss the usage of the pastureland with the locals) without paying rent (outside the city).

Choosing a ger for housing is not only about tradition and culture. It is quite rational in that situation.

It'd be a lot more accurate—not to say more honest—to say the author _estimated_ the number of all the yurts in Mongolia using machine learning. ML algorithms are stochastic; their outputs are whatever the algorithm deems the most probable of the options generated from the given inputs. They barely give a thought to all the ways their count could be wrong—no error analysis, no confidence intervals. There's a meaningless prediction score of 40%, and they blithely add "a hundred or so" to the count.

This is anti-information. People reading this uncritically will come away with completely wrong ideas about the number of yurts in Mongolia, about machine learning algorithms, about data science in general.

Man, I had almost forgotten how much fun it was to read about ML projects of the variety I studied in university, before all the discourse shifted to plain English prompts.

I know of some government entities in Australia doing similar work, but the effectiveness/quality level of the author's work do make me despair for our government a bit. They're blowing years of Very Expensive Consultant spend and they can't even classify an entire parcel of land correctly, let alone count some little yurt-shaped blobs.

"In total I found 172,689 yurts with a prediction score of greater than 40%."

How should one interpet the "prediction score"?

They use a semi-commercial solution (free for educational use).

I'm curious what the topology/architecture of the DL model is like. And are there better ways to approach this problem?

Keeping an eye on the steppe nomads is always a good idea.

Using machine learning, he counted all of the yurts.

Counting all of the yurts that happen to be using machine learning is a way more difficult problem.

No validation versus some kind of ground truth . His training data set is very small and geographically limited. His model is likely pretty inaccurate

I thought the headline was discussing all of the yurts in Mongolia that are using machine learning.

Maybe I missed it but did the article give an answer for the number of yurts?

How much approximately does it cost to rent these amounts of gpu + server?

So hop on the bandwagon before you get left behind by Mongolians in yurts.

All very cool, but wouldn't it be a lot easier to just detect circles within a certain range of size and color using some basic computer vision, like a circle Hough transform?

Where was this when Morrowind came out? :)

What would this have cost?

this is 1 step above yak shaving i suppose

Yurt is a lot of fun to say. Great word.

Nice! Now how will you validate the result?

There are zero yurts in Mongolia using machine learning.

Cool, how can this be used for taxation purposes?

It seems like a waste that you didn't use the 89,259 yurts that are already outlined in OpenStreetMap as input, though you would've probably had issues aligning the outlines with google maps imagery

https://taginfo.geofabrik.de/asia:mongolia/tags/building=ger

I'm also guessing your model doesn't handle yurts that are on the border of a tile.

Finally, that's a much smaller number than I expected for a country of 3 million.

Nice write up, also great to see Docker Swarm being used.

Intrigued by this. What was the rate of false positives? For example are there storage tanks, silos, above ground pools mistaken for yurts?

Nice, thanks for sharing! What would be the best way (and data source) to observe the number of yurts over time?

You estimated. This is not at all a "count."

PSA: Downloading Google Maps satellite imagery tiles is forbidden by the TOS. This is enforced, too, and I'm quite surprised the OP managed to download tiles for all of Mongolia without getting banned.