Nice addition, and good catch on the redundancy of the RMSE in this case! 😄

Two small remarks:

• For this type of statistic I think "count" is the name often used, for example the statistic argument of this SciPy function: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.binned_statistic.html, or the name of this NumPy function: https://numpy.org/doc/2.1/reference/generated/numpy.count_nonzero.html. We could use this here too?
• It might be useful to some users to know the total point count (NaNs included)? If we report a "total count" and "valid count", then users can derive the percent of valid points themselves by dividing the two.

Last thought: I didn't see any function to calculate the valid points in the changes, maybe it's not there yet! In that case I would recommend np.count_nonzero applied to np.isfinite, instead of ~np.isnan, as the latter considers only NaNs but not +/- infinity that are often unusable in stats (and I've had some misadventures with those being propagated in raster data before!).

Good addition!
Two other thoughts:

• I believe your implementation does not exclude masked values. Remember that self.data is a masked array, so invalid pixels are masked instead of being set to NaN. One quick way to get total number of unmasked values is through self.data.compressed().size. Maybe it would be good to add 1-2 tests. For example, the "exploradores_aster_dem" example has data gaps.
• the name of your variable is incorrect. It should not be "percentile" but "percentage".

Regarding @rhugonnet's comment:

It might be useful to some users to know the total point count (NaNs included)? If we report a "total count" and "valid count", then users can derive the percent of valid points themselves by dividing the two.

Not so easy to figure out which ones are the most useful between total count, valid count and fraction of valid pixels. It does not make so much sense to give all 3 as one can derive the 3rd from the 2 others... But I think the percentage of valid pixel is more useful than a count (which is generally gonna be large and not very convenient). Also, total count can be easily derived from the data shape... So I would go for either just the percentage of valid pixels, or percentage + total valid count.

@rhugonnet the function to calculate the valid point was already there, but I used ~np.isnan, then I have changed for np.isfinite following your feedback.

@adehecq the stats are applied on the unmasked data :