Hacker News Reader

Comment by crazygringo | Hacker News Reader

That's bizarre but actually can't be Dark Sky's fault.

Switch to radar view in the app and it will show you the exact geographical distribution of rain in your city right now. This comes from national weather data.

If that's inaccurate then something is wrong with the government data. Or maybe Seattle rain is somehow much patchier at a finer level than government data is able to resolve?

FWIW in New York City its assessment of the next five minutes has always been literally perfect. It might say it's going to go from hard rain to nothing in two minutes... and it does, right on time.

It gets increasingly noisy as you go to 15, 30, and 60 min out, though.

View on HN

Radar data is hugely noisy and it can be difficult to suss out what is cloud cover with no precipitation, or what is precipitation, plus filtering out biomass and dust. All of these show up as radar returns. There can be shadowing. Dark Sky does its own pre-processing/denoising of radar and so that can lead to different results from NOAA.

Ah that's really interesting, thanks for the info.

I'm curious then what users see in Seattle when it's raining out the window but the Dark Sky radar maps says it isn't.

Are there swiss-cheese holes that are obvious processing artifacts? Is the whole city somehow below some threshold so it doesn't show rain anywhere? Are there patterns of extreme local variation in rain that get mistaken for noise and deleted?

The solution is not necessarily simple, but identifying the culprit should be fairly obvious, no?

Really recommend reading http://jackadam.github.io/2011/how-dark-sky-works/ and https://en.wikipedia.org/wiki/Weather_radar#How_a_weather_ra... to get something of a handle on it

For one thing, the resolution of radar is finite where one "pixel" (gate) is quite large, and everything under it is assumed to be constant. Precipitation at a spot on the ground is the result of everything happening from sea level up 40,000', and we don't have exact properties (temperature, dew point, humidity) for the atmosphere at all of those locations. We can fill in some gaps with models on supercomputers, but they're just models.

This was my experience too. Perfect in Boston, recently great in Singapore, pretty much useless in Melbourne.

I’d like to know more about their modeling. How much do they need to specialize on region for accurate prediction or data collection? Are varying prediction qualities based on different climates, or is it from bias in data collection?

Reply on news.ycombinator.com