thomas hopson
The country of Bangladesh (Figure 1) experiences life-threatening floods in the basins of the Ganges and Brahmaputra rivers flowing through the country with tragic regularity. These floods result in loss of life on a scale that often greatly eclipses the deaths due to natural disasters in developed countries. Flooding in these basins can occur on weekly time scales, as occurred during the severe July 2004 Brahmaputra floods that killed approximately 500 people and displaced 30 million over a three week period (Figure 2). And such flooding can occur on seasonal time scales, as occurred during the disastrous floods of 1998 in which the Ganges and Brahmaputra rivers inundated 2/3 of the country for 3 months, killing 1000 people and leaving 40 million homeless. However, with sufficient advance warning much can be done to alleviate the suffering and loss of life and property these floods inflict.
Beginning in 2003 and ongoing to the present, I have been producing operational flood forecasts for this country as part of the USAID-sponsored Climate Forecasting Applications for Bangladesh (CFAB, PI Peter Webster) project at daily to seasonal time-scales (see http://cfab.eas.gatech.edu/shortterm). To produce the most accurate and skillful forecasts that have utility for flood relief and mitigation purposes as far out in advance as possible, accurate numerical 
weather predication forecasts are used in conjunction with satellite-based precipitation products. To meet this end, we have partnered with the European Centre for Medium-Range Weather Forecasts (ECMWF), which operationally provides its 51-member ensemble 1- to 15-day lead-time weather forecasts and 41-member 1- to 6-month lead-time seasonal forecasts to assist with this humanitarian project. We also employ the TRMM and CMORPH precipitation and NOAA CPC-GTS rain gauge estimates to initialize daily boundary conditions and in-stream flow conditions. The fully-automated operational ensemble flood forecasts use these inputs to drive a hydrological multi-model and statistical techniques to remove biases, account for all sources of uncertainty, and enhance forecast skill (Figure 3 a) and b)). These reliable ensemble discharge forecasts then give the probability the river heights will exceed a critical flood stage level (Figure 3 e) and f)). In 2007, this system forecasted two extreme flooding events (July and September). These 1- to 10-day in advance forecasts were utilized by local citizens to evacuate from five highly vulnerable pilot forecast dissemination areas along the Brahmaputra river. These forecasts also assisted them with their daily economic decision-making to optimize their crop yields, protect their fish farm stocks, and their material possessions, etc., along with the well-being of their families (Figure 4).
In addition to producing the operational Brahmaputra and Ganges 1- to 10-day flood forecasts, my work at NCAR has focused on improving the techniques used to produce reliable and skillful ensemble forecasts for a variety of applications, from mesoscale daily temperature forecasts, to the operational Bangladesh flood forecasts just discussed. As well, I have been examining how well dynamically-generated ensemble weather forecasts can forecast their own potential uncertainty. In particular, I have explored the relationship between ensemble spread and forecast skill: what verification measures currently exist to test this relationship, how well do they work and where do they fail, and what other alternatives are there? I have also studied how well a statistical heteroscedastic error model (i.e. error that depends on the magnitude of the forecast variable) derived solely from hindcasts compares with dynamically-generated ensembles in forecasting "forecasting uncertainty". In particular, I argue the real test of the utility of a dynamical ensemble forecast is that it should at least be able to beat this far less expensive simple statistical error model.
