The use of large-scale climate information to predict Central Asia river flows at one- and two-season
leads
Mathew Barlow, AER
Michael Tippett, IRI
Ideas
Central Asia river flows largely driven by snowmelt. Relate river flow to preceding cold season’s snow pack.
Routine snow pack measurements are scarce.
Use local precipitation as a proxy?
Local precipitation influenced by large-scale, potentially predictable, climate variability. Upper level winds
Tropical connection (“Perfect ocean for drought”)
How much information can be extracted from available real-time products?
Where is Central Asia?(For the purposes of this analysis)
Number of years in 3-year period, Nov 1998 - Oct 2001, where precipitation amounts were in the lowest fifth of the yearly values since 1979.
River flow stations in current analysis, 36 years of monthly data, 1950-1985
Previous Results and Background
Precipitation and drought• Agrawala et al. 2001: Recent drought and societal impact for
Central Southwest Asia (CSWA)• Barlow et al. 2002: Drought pattern and west Pacific SSTs + La
Nina; precipitation in eastern Indian Ocean• Hoerling and Kumar 2003: West Pacific SSTs + La Nina give
global drought pattern in model• Tippett et al. 2003, 2005: Role of upper-level winds. West Pacific
model precipitation can be used for CSWA forecasts that have operational skill over recent period.
• Barlow et al. (sub): MJO in E. Indian Ocean affects CSWA daily precipitation; Rodwell-Hoskins hypothesis
River flow• Schär et al. 2004: High predictability of river flow for a Central
Asia river based on ECMWF reanalysis antecedent winter precipitation
Data
•UEA 0.5x0.5 precipitation (consistent with averages from station data)
•NCEP/NCAR reanalysis and CDAS winds, precipitation
•Kaplan SSTs
•24 river flow stations, reporting 93-100% of the time, monthly, 1950-1985, from NCAR ds552.1, ds553.2
No correction for human influence on flows. However, results hold across stations representing a range of flow volumes and elevations, and over full period of record.
•Averaging over bulk of flow for a given year (less sensitive to release timing)•Schar results suggest accounting for human influence increases the strength of the relationship; and relationship continues strongly after dissolution of Soviet Union•Results physically consistent
ClimatologyTopography (km)
Nov-Apr Precipitation “cold season”
May-Oct PrecipitationStation precipitation
Correlation: Wind EOF1/Precip EOF1 = .66
Wind EOF1/NE = .58
First EOFof DJFMprecip
First EOF ofDJFM 200 hPa
reanalysis winds
Winter precipitation related to upper level winds.
Yang et al. 2002
•Negative correlation between Central Asia precipitation and EAJS strength.
•Positive correlation between EAJS strength and Maritime Continent precipitation.
East Asia Jet Stream—tropical connection
Mean Normalized River Flow - Coherence
Central Asia Seasonal Cycle
Seasonal cycles of precipitation, river flow, and vegetation in region
Average Normalized River Flow: Antecedent Winter Precipitation and SSTs
Max Correlation = 0.6Max Correlation = 0.8
Correlation to Antecedent Nov-Mar Precip
Covariance to Antecedent Nov-Mar SST
Repeating correlations separately for each half of the record yields the same patterns, slightly stronger in the recent data
• Local aggregated analysis suggests that local precipitation is a useful predictor of river flow.
• Few precipitation observations in real (or recent) time.
• Follow Schär and try to use analysis precipitation estimates.
• Extract additional information from analysis winds.• Use pattern based regression (CCA).
Canonical Correlation Analysis (CCA)
(Using reanalysis/CDAS wind and model precipitation for ease and consistency in operational use.)
Nov-Mar 200hPa U and model precipitation Apr-Aug River Flows
CCA time series: Nov-Mar U,P (blue) & Apr-Aug River Flows (green); Nino 4 (red)
Average correlation = -0.7
CCA in Prediction Mode1) Compute patterns from CCA on historical data2) Project U,P patterns onto current U,P anomalies to get magnitude3) This magnitude times the river flow pattern is the river flow forecast.
In the current case, essentially predicting the total river flow amount at the start of the season. For cross validation skills, sequentially withhold one season from pattern calculations and forecast for that season.
Cross-Validated Skill Scores
River Flow Stations
Summary
Central Asian river flows can be predicted based on operationally available data, with an average cross-validated skill correlation of 0.43 (including the 3 wayward stations) and 10 stations correlated greater than 0.5.
Considerable spatial coherence in the signal and a relationship to a previously-recognized pattern of large-scale variability.
Possible forecasts refinements include• updating forecast in May and June to account for end of rainy season• using more real-time observed precipitation• targeting individual stations.
Preliminary results suggest that vegetation (NDVI) may be amenable to the same forecasting technique.