Hello,
In working with the C3S Atlas data, I find that across Member models and locations, there are negative and positive values that are exactly -/+ -8.209961. Not only are these very extreme SPEI values, they are also quite different from the second-lowest or largest values.
I suspect these are some lower/upper bound of the SPEI-6 calculations? When one wants to do calculations in SPEI changes etc., would it be best to set these cells to NA, or perhaps adjust them so that they reflect the second-most extreme values present for that particular layer?
I tried looking through the documentation but can’t find particular information on these values.
Dear Merijn,
I need to check this with some experts, meanwhile, please have a look on our SPEI-6 Jupyter notebook, which might provide some useful hints: 1.6. Monthly Standardised Precipitation-Evapotranspiration Index (SPEI) for 6 months cumulation period — C3S Atlas jupyter-book
Thank you, best regards
Andras Horanyi
Hi Andras,
Thank you! I have not been able to find the explanation in the notebook, so if you’re able to retrieve this information that would be most helpful!
All the best
Merijn
Dear Merijn,
Thank you for your patience, I checked this issue with several experts.
The issue is caused by extreme SPI values (see details at [xclim.indices.standardized_precipitation_index() return infinite values · Issue #1416 · Ouranosinc/xclim · GitHub](xclim.indices.standardized_precipitation_index() return infinite values · Issue #1416 · Ouranosinc/xclim · GitHub "Original URL:
xclim.indices.standardized_precipitation_index() return infinite values · Issue #1416 · Ouranosinc/xclim · GitHub
Click to follow link.")). Due to floating-point precision, the Gamma Cumulative Distribution Function (CDF) can return exactly 0 or 1 instead of values very near to 0 and 1, respectively. These 0 and 1 values make the normal inverse (Percent Point Function) return ±∞. The minimum/maximum finite SPI with float64 precision is about ±8.21. Therefore, the simplest solution was applied to provide SPI values to [-8.21, 8.21] in these cases. So in that sense the data values are still real values but using some numerical adjustments.
Please note though that very extreme SPI values should be interpreted cautiously because of statistical uncertainty from the very limited sample sizes.
These values are on the edge of the distribution; most likely the event is wetter/drier than any event during the reference period used to fit the distribution. Therefore, you might consider interpreting both SPI and SPEI only in ranges or categories, such as the ones for instance outlined in the paper ( ERA5–Drought: Global drought indices based on ECMWF reanalysis | Scientific Data ). For example, any value smaller than -3 should simply be interpreted as an extreme drought, and the magnitude of the index should not be analysed.
Thank you, best regards
Andras