How Accurate Is ERA5 Snowfall in Japan? Against 63 Weather Stations, It Reports About Half

Every snowfall figure on this site comes from a weather model. So do most sites that publish worldwide snowfall figures: there is no global network of people measuring snow, so reanalysis data such as ERA5 is the usual source.

It is wrong. Not slightly, and not randomly.

We compared it against 63 Japan Meteorological Agency stations sitting within 15 km of a ski area, over three winters. The model returns 56% of what the ground actually measured.

The gap, station by station

Take the five stations where the model is furthest off:

Station		Measured	Model	Model returns
Takano	Shimane, 570 m	388 cm	103 cm	27%
Daisen	Tottori, 875 m	677 cm	191 cm	28%
Imaoka	Hyogo, 207 m	60 cm	17 cm	28%
Yahata	Hiroshima, 774 m	442 cm	139 cm	31%
Oasa	Hiroshima, 399 m	238 cm	79 cm	33%

Daisen measured 6.8 metres. The model said 1.9. A ski area planning around that figure would be planning for a different mountain. It is also why our own ranking puts Daisen 308th in Japan while the station 500 m away was recording more snow than Hakuba.

Now the five where the model comes out above the ground:

Station		Measured	Model	Model returns
Shiroishi	Miyagi, 86 m	42 cm	65 cm	155%
Kaga-Nakatsuhara	Ishikawa, 65 m	219 cm	245 cm	112%
Iwate-Matsuo	Iwate, 275 m	205 cm	223 cm	109%
Imazu	Shiga, 88 m	114 cm	119 cm	104%
Tonami	Toyama, 69 m	246 cm	251 cm	102%

Look at the elevations. Four of these five are below 100 m, on flat ground, and the fifth is at 275 m. The single closest match, though, is Hakuba at 703 m (447 cm measured, 453 cm modelled over the same three winters), so height alone does not doom a station.

The error is elevation, and it is systematic

Group the 63 stations by height and the pattern is unmistakable:

Station elevation	Stations	Measured (median)	Model (median)	Model returns
0–150 m	18	379 cm	258 cm	73%
150–400 m	21	336 cm	204 cm	55%
400–700 m	16	481 cm	224 cm	47%
Above 700 m	8	569 cm	277 cm	39%

Read the two middle columns across. The ground gets far snowier at the top of the range — 569 cm against 379 cm at the bottom — and the model barely moves: 277 cm against 258 cm. It is not tracking the mountain at all. (The last column is the median of each station’s own ratio, not the ratio of the two medians beside it.)

The rank correlation between a station’s elevation and how well the model does is −0.349. Against the amount of snow itself it is −0.098, which is close to nothing. The model is not failing because there is a lot of snow. It is failing because there is a mountain.

Why a global model cannot see a mountain

ERA5 divides the planet into cells roughly 30 km across and gives each one a single elevation — an average of everything inside it. A cell containing Daisen contains the plain around Daisen too, so the mountain is flattened into a low hill.

Snowfall in Japan comes from cold air crossing the Sea of Japan, picking up moisture, and being forced upward by terrain. Take the terrain out and you take out the mechanism. What the model returns is roughly the snow that would fall on the average of the landscape, and skiing happens on the exception.

That is also why lowland stations do fine. At 69 m, Tonami is the average of its cell.

Where the error is worst, and where it is not

Region	Stations	Model returns
Western Japan (Chugoku, Kinki)	16	44%
Hokkaido	4	50%
Kanto / Chubu	6	48%
Tohoku	24	56%
Hokuriku / Shinetsu	13	77%

Hokuriku and Shinetsu do best because their stations sit on the coastal plain, where the snow arrives without needing a mountain to lift it. Western Japan does worst because its snow falls on isolated peaks — Daisen rises alone out of low country, which is exactly the shape a 30 km cell erases.

What this means for the numbers on this site

Do not read our centimetre figures as measurements. They are model output, and in Japanese mountains they run at roughly half of reality. When we say a resort gets 300 cm, the ground there is probably nearer 550.

Do read the rankings. Every one of the 3,069 ski areas we cover was measured by the identical method, so the ordering survives an error that applies to all of them. Position 1 is genuinely snowier than position 500. The centimetres attached to those positions are not the whole story.

And treat comparisons across terrain with care. A flat, coastal ski area and a high, isolated peak are not being measured equally well by this model — the peak is being shortchanged more. That is the one place where the ranking can mislead, and it is why we publish this page.

For Japan we can check. Outside Japan there is no equivalent network to check against, which is why we do not publish a snow depth for ski areas outside Japan — a number nobody can verify is not a number worth printing. The seasonal snowfall figures we do publish everywhere are for comparing one place with another.

How this was measured

Stations come from the Japan Meteorological Agency, restricted to those within 15 km of a ski area we cover, capped at four per prefecture so that a single snowy prefecture cannot dominate. That leaves 63 stations. Snowfall is the total for the December-to-March period, averaged across the winters of 2023/24, 2024/25 and 2025/26.

The model figure is the ERA5 reanalysis for the same location and the same period, computed exactly as it is for every ski area on this site.

Three limits worth stating plainly. The stations are lower than the ski areas. Weather stations sit in valleys and towns because that is where people are, while lifts run up the slope above them, so the true gap on the piste is likely wider than the one measured here, not narrower. Sixty-three stations is a small sample, and the four in Hokkaido are too few to characterise that island on their own. And a ratio above 100% is not the model being generous — at Shiroishi it reflects 42 cm of snow, where a few centimetres either way moves the percentage a long way.

Sources: Open-Meteo historical weather data, based on the ERA5 reanalysis (Copernicus Climate Change Service); Japan Meteorological Agency, past weather data.