...
Whole table corresponds to 6h rainfall totals for 06-12UTC 28th November 2020
Column 0: Verifying observations
Columns 1-5: Ensemble Mean fields for ECMWF IFS (runs from 00UTC and 12UTC) and COSMO 2.2km (runs from 21UTC only, nominally shown as 00UTC here). PP means post-processed (different approaches for the two systems)
Columns 6, 7: 98th percentiles from ecPoint and post-processed COSMO
Column 8: MISTRAL-style output: 98th Percentile, a blended version of post-processed ENS (12UTC run) and post-processed COSMO (21UTC run same day)
Some points to note from the above:
- Some very localised extremes occurred in this time window (related to flash flood events) - e.g. see pink spots in column 0.
- RAW ECMWF ensemble mean forecasts, at short lead times (as used for the MISTRAL products), were very consistent.
- RAW COSMO ensemble mean forecasts (admittedly only two sets available) were not very consistent. The shorter lead time forecasts showed much more rain in the south of Sardinia, and a bit less in the north.
- Near the east coast the RAW ECMWF ensemble (mean) over-predicted rainfall, whilst COSMO under-predicted; respectively about 30mm and 2mm versus 10-20mm consistently observed.
- ecPoint post-processing of the RAW ECMWF ensemble removed most of the positive bias (based on calibration which tells the post-processing that over-prediction in such situations is commonplace). COSMO post-processing did not remove the negative bias (because bias correction is not built in).
- ecPoint post-processing can in principle shift the spurious coastal maximum inland, to where it is needed, if different bias levels are incorporated via calibration: there are some hints of a shift in this direction. With predictor variable adjustments more could probably be done.
- COSMO mean fields seem to perform better in the NW and SW of Sardinia, where the RAW ECMWF ensemble seems to have insufficient rain.
- 98th percentiles of the post-processed fields show much larger values than the ensemble mean, which is consistent with the observations that show a lot of local variability, with values broadly of the right order (about 2 of the available observations should exceed the values shown).
- Blending can help mitigate against the systematic biases of the two systems particularly when they are "in opposition", as here, although the impact of giving increasing weight to COSMO at shorter leads is also clear to see (column 8).