Ensemble-Based Statistical Prediction of Ethiopian Monthly ...
Transcript of Ensemble-Based Statistical Prediction of Ethiopian Monthly ...
Ensemble-Based Statistical
Prediction of Ethiopian Monthly-to-Seasonal Kiremt Rainfall
1Cooperative Institute for Mesoscale Meteorological Studies/2School of Meteorology
The University of Oklahoma, Norman, OK ([email protected])
Zewdu T. Segele1, Peter J. Lamb1, 2, and Lance M. Leslie2
Kiremt is mostly June-September except in Southwest and West where it starts earlier
Physical & Human Characteristics Population: 85 million (projected to be 174 million by the year 2050 according to the 2010 World Population Data Sheet, Population Reference Bureau , Washington DC)
Area: 1.2 m square km ( ~ twice the size of Texas) Lowest point: -125 m/410 ft below MSL (Danakil depression) Highest point: 4620 m/15157 ft above MSL (Ras Deshen) Blue Nile: contributes more than 85% of the Nile water
Long.
Blue Nile
Eritrea
Djibouti
Somalia
Somalia Kenya
Rainfall Climatology Two main rainy seasons for the northern two-thirds of the country: Feb.-May and Jun.-Sept. (also called Kiremt)
Southwest and West: mono-modal rainfall pattern
Central and Northeast: bi-modal rainfall pattern, one major
East and South: bi-modal rainfall pattern, comparable seasons
Jan Mar May Jul Sep Nov
0
3
6
9
12
15
Jan Mar May Jul Sep Nov
0
3
6
9
12
15
Jan Mar May Jul Sep Nov
0
3
6
9
12
15
Southwest & West (Gore)
East & South (Diredawa)
Central & North (Combolcha)
J a n M a r M a y J u l S e p N o v
0
3
6
9
1 2
Rainfall Pattern for Dry-Summer Areas
Near equatorial southern lowlands are dry during summer Two short seasons; one in spring and another in fall in association with the Inter-tropical Convergence Zone (ITCZ)
These pastoral regions (including Somalia and northern Kenya) are currently experiencing severe drought and famine
Southern lowlands (Negele)
Seasonal prediction is important!
Rainfall Data Distribution
100 rain gauge stations for 1970-99 for all-Ethiopian seasonal prediction
52 stations for 2000-2002
• 11 stations for regional prediction Combolcha for local prediction
Jan Apr Jul Oct Jan Apr Jul Oct Jan Apr Jul Oct-1.0
-0.5
0
0.5
1.0
Month
Cor
relat
ion
Monsoon
Y(-1) Y(0) Y(+1)
West Eq. Pacific East Eq. Pacific
99%95%
99%95%
Correlation between Ethiopian Kiremt Rainfall (1970-99) and Zonal Wind (unfiltered) and SST (filtered at the
ENSO time scale) Concurrent correlations between Kiremt rainfall and zonal wind (> 0.5) and ENSO time-scale filtered sea surface temperature (> 0.9) are very high
Although useful for diagnostics, such concurrent relationships cannot be used for predictions (note below the low SST correlations prior to the onset of Kiremt)
East Pacific
West Pacific
U (1000 hPa)
ENSO (filtered) TIME-LAGGED SST CORRELATION
CORRELATION MAPS
Use both Atmospheric and Oceanic Predictors
SST variations over the Pacific Ocean are widely used for seasonal predictions. However, they explain only a fraction of Kiremt rainfall variability (concurrent unfiltered R2 ~ 35%). Moreover, their effects must be realized through regional circulations => Identify regional circulation predictors also.
Regional: wind, vertical velocity, GPH, temperature, moisture, pressure
Global: sea surface temperature
Time: March, April (pre-Kiremt)
Space: 2.5º lat. x 2.5º lon., 12 vertical levels
Challenge: Too many (> 4000) predictors
Problem of multicollinearity (linearly interrelated predictors)
Predictors To reduce the number of predictors and handle multicollinearity issues,
apply Principal Component Analysis (PCA)
PCA redistributes the variability of a large set of variables into a few components
It produces mutually independent time series
PCA was applied at each level from surface to 100 hPa; up to 470 PCs prescreened as potential predictors
Prediction Design Summary y1 y2 y3 . yi .
yn
Rainfall
Predictor PCs
x11 x12 x13 .
x1i .
x1n
x21 x22 x23 .
x2i .
x2n
xp1 xp2 xp3 .
xpi .
xpn
… … … … … … …
Time
Exclude year i of time series & develop a multiple linear regression equation from the remaining time series by initializing the model with PC1 => Model M1i, Prediction Y1i
Repeat the model fitting procedure by initializing the model with PC2 => Model M2i, Prediction Y2i
Continue until each PC initializes the model, yielding a set of p models and p prediction values for year i => M1i, M2i, …,Mpi, and Y1i, Y2i,…, Ypi
The final prediction for year i is then the average of predicted values of the best models, which are selected based on several criteria that largely involve intra/inter-model correlation strength and significance levels
The above model-fitting procedure is repeated for all years to be predicted
Leave-One-Out Strategy Ensures no information of the future state is leaked into the prediction
Ensemble Procedure
PC1 PC2 PCp
The maximum model ensemble size is the total number of prescreened PCs because each PC initializes a model (the number of unique models, however, is less than the number of PCs as some models may contain identical predictors).
For a leave-one-out prediction strategy, any PC with at least 0.2 correlation magnitude with rainfall (excluding data for predicted year) is used as a potential predictor provided that the correlation is stable (i.e., remain fairly the same) for any 9/10 data-pair segments of the training data
ENSEMBLE SIZE
All-Ethiopian Standardized Kiremt Rainfall Prediction 1990-2002
(leave-one-out, based on predictors observed in April) For the Prediction, only fitted model statistics were used to select ensemble members
For the Analysis, concurrent observations were used to select the best model combination of the predicted Ypi’s
For all (most) years, the ensemble prediction correctly identified the sign ( magnitude) of the observed standardized Kiremt rainfall anomalies.
High correlation predictions (r = +0.82) for Combolcha (northeastern Ethiopia) based on the state of the ocean-atmosphere system observed in April
Deviations from the observations are large for excessively wet (e.g., 1975) or excessively dry (e.g., 1984) years
The interannual variability of monthly rainfall is well captured
Local Prediction of August Total Rainfall Amount 1970-99, Combolcha
(leave-one-out, April predictors)
Key Conclusions • Both regional circulations and global SSTs
are required for skillful Kiremt rainfall predictions
• Ensemble-based statistical prediction technique provides high-quality local and regional forecasts for Ethiopian Kiremt one to two months in advance of onset month
• Leave-one-out (and also retroactive) prediction technique identified observed all-Ethiopian seasonal anomalies
PUBLICATIONS Segele, Z. T., P.J. Lamb, 2005: Characterization and variability of Kiremt rainy
season over Ethiopia. Meteorol. Atmos. Phys., 89, 153-180. Segele, Z. T., L. M. Leslie, and P. J. Lamb, 2009: Evaluation and adaptation of a
regional climate model for the Horn of Africa: Rainfall climatology and interannual variability. Int. J. Climatol., 29, 47-65.
Segele, Z. T., P. J. Lamb, and L. M. Leslie, 2009: Large-scale atmospheric
circulation and global sea surface temperature associations with Horn of Africa June-September rainfall. Int. J. Climatol., 29, 1075-1100.
Segele, Z. T., P. J. Lamb, and L. M. Leslie, 2009: Seasonal-to-interannual variability
of Ethiopia/Horn of Africa monsoon. Part I: Associations of wavelet-filtered large-scale atmospheric circulation and global sea surface temperature. J. Climate, 22, 3396–3421.
Segele, Z. T., P. J. Lamb, and L. M. Leslie, 2011: Seasonal-to-interannual variability
of Ethiopia/Horn of Africa monsoon. Part II: Ensemble-based statistical predictions of Ethiopian monthly-to-seasonal rainfall. J. Climate (In preparation).