Cherry blossom forecast gets a digital aid

Students' research helps predict peak

Vidhya Dass (left) and Elizabeth Brennan used computer modeling to predict when the cherry blossoms would bloom. (Lois Raimondo/Washington Post)

WASHINGTON - Every year Rob DeFeo consults nature to figure out when Washington's cherry blossoms will bloom. The National Park Service chief horticulturalist, who has rough hands and wears a battered leather jacket, studies plants and trees, ponders seasons past, and taps instincts born of decades watching winter turn to spring.

Millions of dollars ride on his forecast, as does as the fate of the city's annual tourist extravaganza, which began yesterday. DeFeo says he has been on target 13 out of 16 years.

But now come Virginia Tech's Vidhya Dass and Elizabeth Brennan, students armed with artificial neural networks, evolutionary computations, the Arrhenius equation, linear regression, and something called fuzzy logic to suggest an alternative to DeFeo's seasoned eye.

Which is to say: Might the brain of a computer some day match human blossom intuition?

It is, if you will, algorithm vs. biorhythm, a finger on the "enter" button vs. a finger in the wind, artificial intelligence vs. a guy who once had 300 species of azaleas in his yard.

The students' idea grew out of an artificial intelligence class they took last spring as part of a master's program at Virginia Tech's Falls Church, Va., campus. Their teacher, assistant professor Chang-Tien Lu, suggested that they try using artificial intelligence to predict the peak bloom period.

The task has traditionally been done by DeFeo, 52, a wry New Jersey native and lifelong horticulturist who is an expert on the life and lore of the renowned cherry blossoms. DeFeo scrutinizes such things as early flowering elms, maples, and cornelian cherry dogwoods, as well as the weather and other recurring clues to the advent of spring.

This year, according to the forecast he issued this month, the peak bloom period would be from March 27 through April 3. He said Friday that yesterday was the peak bloom day, when 70 percent of the blossoms would open. The bloom generally continues for several days beyond the peak period, depending on the weather.

Dass, 33, a native of India, and Brennan, 24, of Baltimore, set out to see whether a computer model might, theoretically, do as well or better, making it easier for tourists to plan visits and officials to plan the National Cherry Blossom Festival.

"We hoped to create a model that would allow the best prediction with the minimum amount of input," Brennan said. The goal was "to see how our artificial techniques compared to human methods." They tried an array of computer methods to see how each worked, the two said in interviews.

Their approach was far different from that of DeFeo, whom they consulted and who admits he understands little of what they report. "I don't have a clue what they're saying," he said Friday.

Neither would most people.

Dass and Brennan said they focused on computational intelligence and essentially tried to mimic the working of the human brain. This involved considering such things as "multiple-layered feed forward neural networks," they wrote in their paper, as well as "delta rule," "topology," and "Stochastic gradient method."

The students started their project in January 2007 and observed the start of the annual blossom bloom. "They were quite striking, very beautiful," said Brennan, who had never seen the flowers before.

The two did not hazard a forecast but plugged in historical data about past blooms and associated weather conditions. Because they used previously recorded data and outcomes, they were able to see which models worked best. They found several models that were accurate to within a few days of past peak dates.

The students say some models, according to their calculations, came three days closer to the peak bloom date than DeFeo's predictions.