Monthly probabilistic drought forecasting using the ECMWF Ensemble system Christophe Lavaysse(1) J. Vogt(1), F. Pappenberger(2) and P. Barbosa(1) (1) European Commission (JRC-IES), Ispra Italy (2) ECMWF, Reading, UK WWOSC, Montr´eal, QC, Canada Joint Research Center ”To provide to policy and decision makers science advice with independent, evidence-based scientific and technical support.” EDO (Europe) http://edo.jrc.ec.europa.eu/edov2 ADO (Africa) http://edo.jrc.ec.europa.eu/ado Introduction Several definitions and forecast products : ① meteorological droughts: based on precipitation ☛ Standardized Precipitation Index (SPI), drought spells ② hydrological droughts: ☛ Palmer Drought severity Index (PDSI), Standardized runoff Index (SRI) ③ agronomical droughts: ☛ Soil moisture anomalies (SMA), Vegetation index ➚ complexity so ➚ uncertainties Objective of this study: Forecasting the meteorological droughts ➔ quantify the predictability of these events ➔ produce an early warning system (1month LT, SPI1) Data and methods Identified adapted products for drought timescales → Ensemble products ENS Forecast : 51 members Hindcast : 5 members Years of hind. : 20 years 1/week Model ”up-to-date” 1*1 resolution Also used : SEAS or S4 Forecast : 51 members Hindcast : 15 members Years of hind. : limited to 20 years 1/month Coarser resol. 1*1 resolution → the unperturbed members (CNTRL) Precipitation (mm) Methodology 10 20 30 40 50 60 Standard Precipitation Index (SPI) 0 60 80 100 0.6 0.8 1.0 Hindcasts CDF(x) Fitting a gamma distribution 40 0.4 2 20 0.2 Ranked accumulated precipitation from the hindcast 0.0 1 0 20 40 60 80 0.8 0.6 CDF(x) 0.4 0.2 Transformation to a normal distribution 0.0 3 1.0 Precipitation (mm) −3 −2 −1 0 1 2 SPI SPI values -1 to -1.5 -1.5 to -2 6 -2 Conditions moderately dry severely dry extremely dry SPI values >2 1.5 to 2 1 to 1.5 Conditions extremely wet severely wet moderately wet 3 Standard Precipitation Index (SPI) V Flexible Rank diagram of the SEAS hindcast model 0.10 • accumulation period • size of domain precipitation from different products / regions V recommended by WMO 0.06 0.04 0.02 precipitation • adapted to compare 0.00 • bias corrected • relative to the antecedent Occurence frequency 0.08 V Interests 1 2 3 4 5 6 7 8 9 10 Sorted members Period of forecasts : Nov. 2012 - Nov. 2013 Period of hindcasts : Nov. 1992 - Nov. 2012 11 12 13 14 15 16 Standard Precipitation Index (SPI) Climatology of the SPI in Europe Droughts observed during the period [1992-2012] SPI < -1 SPI < -1.5 SPI < -2 for SPI < -2, database too short Ensemble scores General evaluation of the SPI (median) • better score for ENS than SEAS • sensible to small variations of SPI • strong dependency upon the meteorological situation → large temporal variability Ensemble scores Distribution of the individual members for the extreme SPI 1 2 3 Floods 0 1 2 0 Ratio 4 5 SPI observed -2 Drought -1 -2 -1.5 -1 -0.5 Drought 0 0.5 1 1.5 SPI forecasted 2 Floods Ensemble scores Distribution of the members Mean and spread of members 3 1.5 Distribution of ranked members Mean SPI forecasted SD SPI forecasted 0.0 0.5 ensMeanTot 0 −1 −2 −0.5 −3 SPI 1 1.0 2 SPI observed < −1.5 SPI observed > 1.5 0 10 20 30 Ranked members 40 50 −2 −1 0 SPI observed 1 2 Ensemble scores Obs = ’binary’ Forecast a drought: 1.0 Pred = ’probability’ • slightly improvement of 0.6 Hit rate 0.4 number of members and events ENS ~ 0.71 SEAS ~ 0.67 0.2 • very sensible to the 0.8 ROC score 0.0 the ENS 0.0 0.2 0.4 0.6 False alarme rate 0.8 1.0 1.0 0.4 0.0 0.4 0.2 ENSEMBLE 0.0 • artificial 0.2 0.8 0.0 0.6 0.2 0.8 0.6 0.4 SEAS 0.0 • comparable results 0.2 Forecast a drought: Reliability diagrams 0.4 1.0 Ensemble scores 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 Forecast probability, y 1.0 0.4 0.4 0.0 0.0 0.6 0.2 0.2 0.8 0.4 0.6 0.6 1.0 1.0 0.8 0.6 0.4 SEAS 0.2 ENSEMBLE 0.0 0.0 for SPI < -2 0.8 i 0.2 • no significant signal 0.6 Forecast probability, y i improvment with diff thresholds 0.4 0.0 0.2 0.4 0.6 Forecast probability, yi 0.8 1.0 0.0 0.2 • Significant reliability for SPI < -1 and -1.5 • No information about the missed events. • Brier better for SPI : 0.14 reliability > resolution 0.4 0.6 i 0.8 1.0 Spatial and temporal variabilities • Mean = 0.71 • Spatial variability of about 10% of the ROC • Larger variability depending upon the weather conditions → No significant seasonal variation of the ENS performance Developing a warning system Provide to stakeholders a clear and robust information for a warning system Obs and pred = ’binary’ ➔ Identification of products : • • • • individual values (quantiles, median) integrative (mean) probabilistic (pourcentage of the members) deviations / differences between members • consistency of the members • provide an assessment of the uncertainty ➔ Definition of thresholds (same nb. of events / obs.) What is the accuracy of a drought warning ? Is there an index better than the others ? Evaluation of the index SEAS 0.6 0.4 0.5 POD / FAR 0.6 0.5 0.3 0.4 0.2 0.3 0.2 POD / FAR 0.7 0.7 0.8 0.8 0.9 0.9 ENS Q13 Q23 MED Q77 Q88 SpL Spl SpD SpF Mean Q13 Q23 Methods MED Q77 Q88 SpL Spl Methods • 1/3 of droughts detected one month in advance • ± 7% following the method used • Larger variability of SEAS using these criteria SpD SpF Mean Evaluation of the index 0.25 0.20 0.15 0.10 Extreme Dependency Score 0.00 0.05 0.25 0.20 0.15 0.10 0.05 0.00 Extreme Dependency Score 0.30 SEAS 0.30 ENS Q13 Q23 MED Q77 Q88 SpL Spl SpD SpF Mean Q13 Q23 MED Methods Q77 Q88 SpL Spl SpD Methods • More adapted for extreme values (Stephanson et al. 2008) • larger variability between methods • Behaviour of the driest members more related to the droughts observed SpF Mean Evaluation of the index 0.10 0.05 Equitable Threat Score 0.8 0.6 0.4 0.00 0.2 0.0 POD vs Percent correct 1.0 0.15 Using the percentage of members 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 Percent • Opposite POD / PC evolutions • Optimum ETS around 30% of members • Similar results with SEAS 0.5 0.6 Percent 0.7 0.8 0.9 1 Evaluation of the index −1 −2 SPI 0 1 Uncertainty based on the SPI values/distribution ? −3 Hits False alarms Misses Correct neg. Q13 Q23 MED Q77 Members • SPI lower for misses in rel. to correct neg. • No distinction for hits and false alarms Q88 Evaluation of the index Uncertainty based on the spread of the members : Drought forecasted yes no Drought observed yes no 2.31 ± 0.4 2.37 ± 0.4 1.99 ± 0.4 1.88 ± 0.3 ➔ No significant differences ➔ Difficult to assess the uncertainties Conclusions/Perspectives X 1/3 droughts correctly detected X 10% of variation following the method used but main errors from the model X ENS slightly better than SEAS X No clear indicators to provide the uncertainties ➔ sensitivity to the horizontal resolution (ongoing) Next step : • droughts prediction using predictors (WT, large scale components over Europe and Africa) → improvement of the warning system ? • Dry spells (short term drought events using 10d accumulation periods) → especially important over region with low resilience