March 18, 2009 | NCAR scientists are exploring the use of a novel statistical technique to help steer intercontinental flights away from thunderstorms (convection). Over the North Atlantic and North Pacific, air traffic controllers have little weather data: satellite images are generated only once every three hours, and surface radars cannot track storms far at sea.
A NASA-funded project led by Cathy Kessinger is developing 1- to 2-hour forecasts (nowcasts) of oceanic convection. The team uses several techniques based on satellite data to produce “interest values” that correspond to the likelihood of convection. The near-future positions of storms, based on their recent history, are represented as polygons. Although this nowcasting technique is computationally efficient, the polygons omit finer-scale detail.
Huaqing Cai is exploring the use of random forests to improve the nowcasts. The forests are made up of many decision trees, each with its own mix of prediction variables (such as satellite-derived cloud height or type) that are randomly drawn from a larger set. Each decision tree casts a yes-or-no “vote” on the expected interest value at each point in space and time. By pooling the votes, scientists can gauge the likely extent of convection.
In one RAL test, a random forest of 200 decision trees produced one-hour nowcasts for the region around 2007’s Hurricane Dean. When the vote tally at each point is color-coded, the resulting graphic looks much like a satellite image. In the case of Dean, the nowcast closely matches convection observed by NASA’s Tropical Rainfall Measuring Mission satellite.
“The random forest technique tends to give you more realistic storms,” Huaqing says. He adds that the technique should have an increasing edge over its rivals when carried out for longer time periods, because the random forests can easily incorporate weather variables that shape storm development over time.