Accuracy of Long-Range Probability Flood Forecasts for Iowa Rivers

Project Personnel (University of Iowa): Dr. Allen Bradley (Professor), Dr. Anton Kruger (Associate Professor), and Mohammed Habib (Graduate Student).

Introduction
After the 1993 Upper Mississippi River Flood, the National Weather Service (NWS) began implementing its Advanced Hydrologic Prediction Services (AHPS), choosing the Des Moines River as its first demonstration basin.  One component of AHPS uses ensemble streamflow prediction (ESP) techniques to issue forecasts of the probability of flooding in the upcoming season (the next 90 days). Our project intends to estimate the accuracy of these long-range flood forecasts, or in other words, how likely is the river to flood once it has been forecasted?

Figure 1. National Weather Service (NWS) locations in Iowa (squares) where seasonal river forecasts are issued by the North Central River Forecast Center (NCRFC). NCRFC forecast groups in Iowa are the Des Moines River (DES), the Cedar-Iowa Rivers (CIW), other tributaries in Iowa (TIA), and along the main stream of the Mississippi River (M10 and M19). Forecast locations with more than 30 years of historical stream gage records (purple squares) will be used to evaluate the quality of long-range flood forecasts.

Study Area
Operational long-range river forecasts are made for the eastern two-thirds of Iowa by the NWS North Central River Forecast Center (NCRFC) in Chanhassen, Minnesota (Figure 1).

Forecasters use recent weather observations to assess current moisture conditions within a river basin, and then run computer models for different possible scenarios.  For instance, what would the river flows be over the next 90 days if the weather was the same as observed last year?  These “what if” scenarios can be run for any year where historical weather observations are available (in practice, about 50 years into the past).  By analyzing the scenarios with statistical methods, forecasters can predict whether a flood is more (or less) likely given the current moisture conditions in the basin (Figure 2).  

Figure 2. An example of an operational long-range flood forecast issued by the NWS on 19 Jan 2009 for the Iowa River at Marengo. The forecast shows the probability of different flow levels in the upcoming season (90 days). The probabilistic forecast (black triangle) can be compared to the forecast model’s historical simulation of past flows over the season (blue circles). For instance, flows in excess of the flood level typically occur about 63% of the time (from late-January to late-March). However, the forecast for the 2009 season calls for a 97% chance of a flood, a significantly elevated risk of flooding. The Iowa River at Marengo did in fact exceed the flood level in February 2009.

To evaluate the accuracy of these long-range forecasts, we can compare the forecast probability of a flood with what actually happens.  If a flood is forecasted to be more likely, one would expect that the river would flood more often (at least more often than it does on average).  Such an assessment can only be done by comparing the forecast probability with the outcomes issued over many years (Figure 3).  Of the 133 NCRFC forecast sites in Iowa, 54 have at least thirty years of flood observations to do such analysis (see Figure 1).  Our evaluation of the quality of long-range flood forecasts will focus on these sites.

Figure 3: Past performance of seasonal flood forecasts (issued on 23 Jan) for the Iowa River at Marengo for 1957-1999. The probability of a flood (f) and the observed maximum daily flow in the subsequent season are shown (blue circles). If the forecast is that a flood is more likely (f > 0.63) and a flood occurred (a flow above the flood level), the forecast is skillful (shaded area). Likewise, if the forecast is that a flood is less likely (f < 0.63) and no flood occurred, the forecast is also skillful. Note that when a probability of 0.97 was forecasted in the past (the forecast probability of a flood in 2009, as shown in Figure 2), a flood did occur during the season. These historical results and verification measures provide insights on the performance of NWS long-range flood forecast in Iowa.

Methodology
Using the operational forecast model from the NWS North Central River Forecast Center (NCRFC), we will generate past forecasts retrospectively, for a period from 1950 to 1999 (the historical period available for operations).  We will compare what would have been forecasted historically had AHPS been available, with what happened when forecast flood probabilities were unusually high (or low).  The comparison will provide insights on the skill of long-range forecasts, how the skill varies throughout the Iowa flood season, and the potential value of long-range forecasting for flood planning and advance preparation.