(more details : francis.pouponneau++at++meteo.fr or samuel.westrelin@ meteo.fr)
The scores (bias and root mean square error) plotted
are calculated against SYNOP over the whole domain ALADIN-FRANCE. Scores are
averaged over the first semester 2001.
They use about 200 synoptic stations over the domain.
The RMS rises with forecast range from 0.6 to 2.3 hPa. The model underestimates this parameter from 12 hours range till 48 hours to reach -1. hPa.
We can see a diurnal cycle of the bias (overestimated at 12 UTC validity time - forecast ranges 12 and 36 hours - at maximum 0.5°K and underestimated during night at maximum -0.7°K) which is on average -0.5°K.
Its RMS is quite stable with ranges. It is not biased on night but clearly underestimated on afternoon. The mean bias over ranges is about -0.3 m/s.
2. COMMENTS OF CONTINGENCY TABLE - FIRST SEMESTER 2001
Tables are the cloud cover and precipitation contingency tables for ALADIN-FRANCE 36 hours forecast range. Units are octas for cloud cover and mm/6 hours for precipitations.
We present several contingency indexes :
-pnd : probability of no detection
-far : false alarm rate
-pod : probability of detection
-fbi : frequency bias index (the closer to 1 the better)
-ts : threat score (the closer to 1 the better)
-tss : true skill statistics (the closer to 1 the better)
They are computed for different thresholds. It gives the model skill to detect phenomenons over each threshold. For example, considering the threshold 3-6 for cloud cover, a two by two contingency table is rebuilt with one class mixing classes 3-6 and 7-8 (phenomenons whose intensity is equal or higher than the threshold) and another class being the 0-2 class. Indexes are computed over this new table.
For threshold "3-6", pod is 84% and far very low, 8%. But with increasing threshold, pod drops to 42% (less than one event over two is detected) and far rises to 22%.
Threshold "0.1-2" measures the model ability to discriminate the rain/no rain events. far is high (when rain is forecast, on 44 cases over 100 it does not rain) but this is at least partly due to a bias in the methodology (rain gauge measurement compared to the nearest model grid point) because we compare what is representative of the model scale (mesh size, about 100 square km) to a rain gauge (representative of a few square meters). pod is 72%.
Considering increasing thresholds, which correspond to more severe meteorological events, the quality of the indexes falls dramatically. For "10 and more" threshold, far is 79% and pod, very poor, 22%.
Note that the model can be climatologically well tuned (fbi very close to 1 for "2-10" and "10-more" thresholds) but it does not mean it gives the right amount of precipitations at the right place and time.
|
Home |