Frequency of precipitation

[joelfiddes]

Problem: precipitation frequency too high

Probable reason: comes from the fact ( I think) that we do a 2d interpolation which will lead to more non zero events (even if very small) than just a single gridbox would contain. This can of course cause problems in terms of frequency related processes.

Solution:
(1) filter small precip events out (lose precip, conserve frequency)
(2) allocate small precip amounts to larger wet days (conserve annual total and frequency) - take frequency from the gridbox? Just apply a simple threshold of 1mm/day
(3) others?

Hi Joel,

thanks a lot for the data!

Quick analysis - toposcale temperature looks better than original era5, R^2 vs AWS measurements is respectively 0.87 and 0.81.

For precipitation, the toposcale predicts at least some precipitation 4.5x more frequently compared to era5 - 82 % of the hourly timesteps have some precipitation, vs 18 % for era5 (see attached plot). At daily aggregation (used by the mass balance model) toposcale has some precipitation on 96 % of days, vs 46 % for era5. This is not a problem for the model, but I'm just curious - have you observed a similar pattern at other locations?

Cheers,

Enrico

[ArcticSnow]

It'd be good plotting the 9 grid cell cumulative precipitation to see if this is from an interpolation artefact or very different adjacent gridcell. This is a significant amount of precip.

[joelfiddes]

this is the annual total for 2020, x mark the tscale point (this is Pamirs, Tj by the way):

[joelfiddes]

somewhere 250-350mm in the 3x3, so not too extreme a gradient. By definition, interpolation is going to increase frequency though, right? As you are more likely to have precip in at least 1 of the 9 cells than the centre cell.

[krisaalstad]

I think this is simply due to the inverse distance weighting approach using many neighbors (which I would recommend for speed and smoothness). If you have a large domain (lots of ERA5 grid cells) and use many neighbors, then you can have some drizzle everywhere in your domain as long as it's raining for one of the ERA5 grid cells. So it won't add much precip at all, but surely aritifically increases the frequency. The simple solution here is to just apply a threshold as Joel suggests, although I think 1 mm/day is too strict. I would recommend a much lower threshold, like 0.1 or even 0.01 mm/day since these artificial drizzle numbers should be quite small given the large distances involved in the interpolation.

[ArcticSnow]

I am running in the same issue. I have precip so often in Finse. I checked the code, and for the precipitation we do the IDW interpolation horizontally. If I pick one of the ERA5 cell in the vicinity, I obtained basically the same signal so our interpolation does not add precip. The problem is I think coming from ERA5 data itself.