Seasonal forecasting of East African Rainfall

1
Seasonal forecasting of East African Rainfall

• Duncan Ackerly
• David Grimes
• Emily Black
• Dept of Meteorology
• University of Reading

2
Overview of Talk

• Background
• Need for seasonal forecasts
• East African rainfall
• Methodology
• local forecast
• statistical regression model based on SSTs
• Results
• assessment of skill
• comparison with persistence, climatology and
  random forecast

3
Background

• The majority of people in East Africa rely on
  subsistence agriculture
• Agriculture is rainfall limited
• Accurate seasonal forecast would allow
• better economic planning
• earlier identification of food shortage

4
Case study

• Use March April May (MAM) rainfall in Kenya as
  case study

5
Approaches to seasonal forecasting

• Two approaches to seasonal forecast
• NWP model
• use numerical weather model run out to time scale
  of months
• statistical
• look for statistical relationship between
  rainfall and likely predictors

6
This project

• Statistical model
• Monthly mean Sea Surface Temperature (SST) with
  lead time of 1 or more months as predictor
  variables
• rainfall anomaly within local zones as target
  variable

7
Methodology

• Divide Kenya into zones based on spatial
  correlation of MAM rainfall.
• For each zone
• identify SST regions in preceding months which
  best correlate with MAM rainfall anomaly
• Perform regression to establish relationship
  between SSTs and rainfall anomaly
• Use cross validation to test robustness of method

8
Raingauge data

• 53 gauges from 1993 to 2001 used to establish
  zones
• 27 gauges from 1949 to 1985 used to calibrate and
  validate regression model for each zone

9
Rainfall zones

• Rainfall zones were defined as areas delineating
  clusters of highly correlated MAM gauge totals

2
4
2
4
5
6
1
6
1
3
3
5
10
SST teleconnections

• Correlation between mean MAM for Zone 1 and SST
  using NCEP reanalysis data

11
SST teleconnections
2
5
4
6
1
3
12
Cross validation time series comparison
13
Cumulative decile forecast results
14
Forecast Success by Quintile

• 1very low 4 high
• 2 low 5 very high
• 3 average

15
Conclusions

• A local seasonal rainfall forecast model based on
  correlation with SSTs with a lead time between 1
  and 3 months has been developed
• The model has been tested using MAM rainfall data
  from Kenya
• Zones closest to the coast show highest lagged
  correlations with Indian Ocean SSTs.
• Inland zones show highest lagged correlation with
  areas in the Atlantic and Pacific
• For 4 out of 6 zones there is significant skill
  in forecasting rainfall anomalies to within two
  deciles

16
Intra-zonal correlation matrix