[Questions] Is it possible to implement diurnal MBCn on hourly data with xclim sdba?

[DamienIrving]

Setup Information

• Xclim version: latest

Context

I'm interested in whether it's possible to use the sdba module for the bias correction of hourly timescale data, taking into account the diurnal cycle?

It is well known that biases are not necessarily constant throughout the seasonal cycle (e.g., autumn biases may differ from spring biases), which is why the sdba module grouping functionaity allows for different correction factors to be computed for each month (or the use of a moving window across a calendar year). When working with daily timescale data, that's the only time grouping that's typically needed.

When working with hourly data, the fact that biases are not necessarily constant throughout the diurnal cycle (e.g., daylight biases may differ from nighttime biases) also becomes relevant. Recent studies have used a diurnal version of the MBCn method, where "bias corrections are computed for each hour, using a 3 hour moving window to pool all hourly values within a given month before using the MBCn algorithm" (see Faghih et al 2022 for the method description, which is also applied by Weeding et al 2024). In other words, they apply both seasonal and diurnal grouping.

Is it possible to apply dual groupings like this with the sdba module as it is currently configured, or would that require changes/updates to the module?

Steps To Reproduce

No response

[coxipi]

Hi Damien,

Ben Weeding was also wondering this in another discussion : #1894 .

The short answer is no, MBCn doesn't support this. But, the old implementation was more flexible, and apparently something could be worked out, maybe the other discussion thread will give you ideas.

I think the main problem with the new implementation would be that there is no interpolation between group values right now. In regular sdba methods for daily dataset, if you perform a monthly or seasonal adjustment, there will be a form of interpolation along the monthly/seasonal direction. In other words, you get an adjustment factor for the month of January, which is directly applicable to jan 15th, an adjustment factor for the month of February, which is directly applicable on Feb 14, Feb 15. For the days in-between Jan 15th and Feb 14-15th, this should be an interpolation of the two adjustment factors.

The way MBCn is currently coded is aimed to optimize the dayofyear computation. The way the grouping is done is less flexible in some aspects, and more in others. Instead of using sdba.Grouper and predefined frequencies understood by Pandas, it constructs an array of time indices corresponding to a given group, a bit like what is done in numpy groupies https://pypi.org/project/numpy-groupies/ . This makes it harder (I think) to implement the interpolation.

However, I said it was more flexible in some regard. Being agnostic on the Pandas frequencies, I think it would be reasonable to construct your specific double time grouping. Currently, MBCn only accepts a group argument that is highly constrained, then those time indices inspired by numpy groupies are built. I think it could be interesting to let users give those time indices instead, which would let you build any grouping your imagination can come up with.

There would always be the caveat that there is no interpolation on the group values, as I discussed above, so I wonder if that would still be of interest. I looked very fast at the 2022 paper, and I didn't see interpolation mentioned in this context, but perhaps I should read more carefully.