BOSTON—When the Red River began to rise in the spring of 1997, the National Weather Service (NWS) predicted that the water would peak at 15 meters in Grand Forks, North Dakota. Unfortunately, the river crested 1.5 meters higher and caused more than $1 billion of damage. Part of the problem was that NWS did not communicate the uncertainty in its estimate. City officials have said that had they known the possible range, they might have been able to protect some areas from flooding.
It was reading a scientific article about these problems with the Red River predictions that motivated Frauke Hoss, a civil engineer, to study the error inherent in such forecasts. Hoss, who is a doctoral student at Carnegie Mellon University, in Pittsburgh, Pennsylvania, has developed a new method for communicating this uncertainty to emergency responders, and she presented it here yesterday at the annual meeting of AAAS (which publishes ScienceNOW). "This is great," says Ben Orlove of Columbia University's Center for Research on Environmental Decisions. "It's more clear."
NWS issues river forecasts based on analysis from 13 centers around the United States. For longer-term outlooks, which assess the risk of flooding over 3 months, the centers use a probabilistic method that provides a calculation of uncertainty. But daily river forecasts are done another way, without this estimate. It's clear that the short-term forecasts can be imprecise when they look beyond 1 day, sometimes missing the mark by several feet and tending to under predict water levels. These uncertainties are inherent in the hydrologic model used by WS and have been difficult to minimize. Hoss said she doesn't know why NWS doesn't include uncertainties in its daily river forecasts, such as a percent error or a range of possible river heights. "Anything would help!"
Hoss decided to see if she could do it herself. She had a hard time finding archived forecast data in a useable format, but she eventually got records for the Neosho River in Miami, Oklahoma, for 2007 through 2011. Hoss then took the forecasts that had been made on 1524 days and compared them to seven factors that influence river height, such as upstream water levels and rainfall intensity. By applying a machine learning tool (a statistical technique often used for programming robots), Hoss could investigate how various combinations of these factors influence the probabilities of a flood on the Neosho River. This means that on any given day, a simple calculation of these seven variables—which are all available online in real time—will reveal the likelihood that the river forecast will in fact correctly predict flooding 48 hours in advance.
The algorithm divided the data into six types of situations. For four of these, the forecast was more than 97% likely to be correct in predicting a flood. But there were two situations, which occurred on 41 days, in which it was difficult to know the reliability of the forecast. "What makes me happy is that the deep uncertainty is limited to a few combinations of conditions," Hoss said. She emphasized that the 24-hour predictions from the NWS are very accurate, but this tool could help emergency managers know when to have more confidence in using the 48-hour forecasts, perhaps giving them more time to protect infrastructure. Hoss thinks the approach could also be used to provide guidance for barge operators, who must know river heights in order to figure out how much cargo they can load.