Operational Flood Forecasting for Bangladesh:

Tom Hopson, NCARPeter Webster, GTA. R. Subbiah and R. Selvaraju, ADPC

Climate Forecast Applications for Bangladesh (CFAB): USAID/CARE/ECMWF/NASA/NOAA

Bangladesh Stakeholders: Bangladesh Meteorological Department, Flood Forecasting and Warning Center, Bangladesh Water Development Board, Department of Agriculture Extension, Disaster Management Bureau, Institute of Water Modeling, Center for Environmental and Geographic Information Services, CARE-Bangladesh Contact: hopson@ucar.edu

Overview:Bangladesh flood forecastingCFAB forecasting context

II.Forecasting techniques -- using Quantile Regression for:1. precipitation forecast calibration2. Post-processing to account for all errors

III. 2007 Floods and Warning System Pilot Areas

(World Food Program)Damaging Floods:large peak or extended durationAffect agriculture: early floods in May, late floods in September Recent severe flooding: 1974, 1987, 1988, 1997, 1998, 2000, 2004, and 20071998: 60% of country inundated for 3 months, 1000 killed, 40 million homeless, 10-20% total food production2004: Brahmaputra floods killed 500 people, displaced 30 million, 40% of capitol city Dhaka under water2007: Brahmaputra floods displaced over 20 million River Flooding

CFAB Project: Improve flood warning lead time Problems:Limited warning of upstream river discharges (Dhaka: ~24-48 hr warning, only)Precip forecasting in tropics difficultSkillful CFAB forecasts benefit from:1. Large catchments => river discharge results from integrated inputs over large spatial and temporal scales2. Skillful data inputs: ECMWF, TRMM, CMORPH, CPC-rain gauge3. Partnership with Bangladeshs Flood Forecasting Warning Centre (FFWC)=> daily border river stage readings useful for data assimilation

Daily Operational Flood Forecasting Sequence

Statistically corrected downscaled forecasts

Generate forecasts

Update soil moisture states and in-stream flows

Calibrate model

Generate hindcasts

Generate forecasts

Distributed Model Hindcast/Forecast Discharge Generation

Generate hindcasts

Generate forecasts

Above-critical-level forecast probabilities transferred to Bangladesh

Convolve multi-model forecast PDF with model error PDF

Generate forecasted model error PDF

Generate hindcasts

Generate forecasts

Updated outlet discharge estimates

Calibrate multi-model

Updated distributed model parameters

Calibrate AR error model

Multi-Model Hindcast/Forecast Discharge Generation

Updated TRMM-CMORPH-CPC precipitation estimates

Forecast Trigger:

ECMWF forecast files

Lumped Model Hindcast/Forecast Discharge Generation

Discharge Forecast PDF Generation

Generate hindcasts

*

Example of Quantile Regression (QR)Our application

Fitting precipitation quantiles using QR conditioned on:Reforecast ens

ensemble mean

ensemble median

4) ensemble stdev

5) Persistence

*

Calibration Procedure

For each quantile:

Perform a climatological fit to the data

Starting with full regressor set, iteratively select best subset using forward step-wise cross-validation

Fitting done using QRSelection done by:Minimizing QR cost functionSatisfying the binomial distribution

3) 2nd pass: segregate forecasts into differing ranges of ensemble dispersion, and refit models.=> have different calibration for different atmospheric stability regimes=> ensures ensemble skill-spread has utility ProbabilityPrecipitationobsForecastPDFPrecipTimeForecastsobservedRegressors for each quantile: 1) corresponding ensemble 2) ens mean 3) ens median 4) ens stdev 5) persistence

Significance of Weather Forecast Uncertainty on Discharge Forecasts 3 day4 dayCalibrated Precipitation ForecastsDischarge Forecasts1 day4 day7 day10 day1 day4 day7 day10 day

Daily Operational Flood Forecasting Sequence

Statistically corrected downscaled forecasts

Generate forecasts

Update soil moisture states and in-stream flows

Calibrate model

Generate hindcasts

Generate forecasts

Distributed Model Hindcast/Forecast Discharge Generation

Generate hindcasts

Generate forecasts

Above-critical-level forecast probabilities transferred to Bangladesh

Convolve multi-model forecast PDF with model error PDF

Generate forecasted model error PDF

Generate hindcasts

Generate forecasts

Updated outlet discharge estimates

Calibrate multi-model

Updated distributed model parameters

Calibrate AR error model

Multi-Model Hindcast/Forecast Discharge Generation

Updated TRMM-CMORPH-CPC precipitation estimates

Forecast Trigger:

ECMWF forecast files

Lumped Model Hindcast/Forecast Discharge Generation

Discharge Forecast PDF Generation

Generate hindcasts

Producing a Reliable Probabilistic Discharge Forecast

2004 Brahmaputra Ensemble Forecasts and Danger Level Probabilities3 day4 day5 day3 day4 day5 day7 day8 day9 day10 day7-10 day Ensemble Forecasts7 day8 day9 day10 day7-10 day Danger Levels

Brahmaputra Discharge Forecast VerificationRank HistogramsBrier Skill ScoresCRPS Scores

Daily Operational Flood Forecasting Sequence

Statistically corrected downscaled forecasts

Generate forecasts

Update soil moisture states and in-stream flows

Calibrate model

Generate hindcasts

Generate forecasts

Distributed Model Hindcast/Forecast Discharge Generation

Generate hindcasts

Generate forecasts

Above-critical-level forecast probabilities transferred to Bangladesh

Convolve multi-model forecast PDF with model error PDF

Generate forecasted model error PDF

Generate hindcasts

Generate forecasts

Updated outlet discharge estimates

Calibrate multi-model

Updated distributed model parameters

Calibrate AR error model

Multi-Model Hindcast/Forecast Discharge Generation

Updated TRMM-CMORPH-CPC precipitation estimates

Forecast Trigger:

ECMWF forecast files

Lumped Model Hindcast/Forecast Discharge Generation

Discharge Forecast PDF Generation

Generate hindcasts

Five Pilot Sites chosen in 2006 consultation workshops based on biophysical, social criteria:Rajpur Union -- 16 sq km-- 16,000 pop.Uria Union-- 23 sq km-- 14,000 pop.Kaijuri Union-- 45 sq km-- 53,000 pop.Gazirtek Union-- 32 sq km-- 23,000 pop.Bhekra Union-- 11 sq km-- 9,000 pop.

2007 Brahmaputra Ensemble Forecasts and Danger Level Probabilities7-10 day Ensemble Forecasts7-10 day Danger Levels7 day8 day9 day10 day7 day8 day9 day10 day

Community level responses to 2007 flood forecasts

Planned evacuations to identified high grounds with adequate communication and sanitation facilities

Economically, they were also able to:Move livestock to high lands with additional dry fodder.

Early harvesting of rice and jute anticipating floods.

Protected fisheries by putting nets in advanceSelvaraju (ADPC-UNFAO)

Conclusions

2003: Daily operational probabilistic discharge forecasts experimentally disseminated based on lumped model and 51-member ECMWF ensemble

2004: -- Multi-model and post-processing approach operational-- initializing watersheds using TRMM / CMORPH-- Forecasts automated-- CFAB became Bangladesh federal government entity-- forecast the severe Brahmaputra floods

2005: CFAB became HEPEX test bed

2006: -- Forecasts incorporated into national flood warning program and hydraulic model-- 5 vulnerable pilot areas designated and trained on using 1-10day probabilistic forecasts.

2007: 5 pilot areas warned many days in-advance during two severe flooding events

2008-2009: Ongoing expansion of the warning system thoughout Bangladesh

Further technological improvements through HEPEX test bed collaborations

Thank You!hopson@ucar.edu

*Focus of this talk: re-acquaint the HEPEX community to the Climate Forecast Applications for Bangladesh (CFAB) project (which became a HEPEX test bed in 2005) and encourage technical collaborations with the research community. Forecasting scheme is modular enough to incorporate other techniques leading to comparison studies and forecast skill improvements.

CFAB: Project began in 1999, became operational in 2003, becoming a Bangladesh Govt entity in 2005, with a program to directly disseminate the forecasts down to the household level beginning in 2006. To our knowledge, first time USAID has funded a university-based research group to develop an operational humanitarian assistance warning program.

Affiliations:Hopson NCAR/RALWebster -- Georgia Institute of Technology, Principal InvestigatorSubbiah and Selvaraju -- Asian Disaster Preparedness Centre (non-profit based in Bangkok)

Special acknowledgments:Funding Agencies: USAID-Dhaka and CARE-Bangladesh; ECMWF for providing their ensemble forecasts free-of-charge (!!) to support this humanitarian effort.NASA-Goddard for providing their satellite-based TRMM rainfall estimation product (Tropical Rainfall Measuring Mission)NOAA-Climate Prediction Center for their CMORPH rainfall estimation product(weve been using the TRMM and CMORPH precipitation estimates and rain gauge interpolated fields operationally for 5 years, I.e. 2004)Stakeholders w/in Bangladesh -- no need to list, just acknowledge active Bangladesh institutional involvement/stackholding.first part of talk give some context to the CFAB forecasting project.

second part will briefly cover how we use Quantile Regression to calibrate: 1) the EPS pre