Wednesday, July 3, 2019

Artificial Intelligence Predicts Lightning Strikes before the Flash


Predicting where lightning will strike. Artificial intelligence can do that accurately.

At the beginning of June, the German Weather Service counted 177,000 lightnings in the night sky within a few days. The natural spectacle had consequences: Storm gusts, hail and rain injured several people. Together with the German Weather Service, computer science professor Jens Dittrich and his doctoral student Christian Schön from the University of the Saarland are now working on a system that will predict local thunderstorms more precisely than previously. It is based on satellite imagery and Artificial Intelligence. In order to investigate this approach in more detail, the researchers receive € 270,000 from the Federal Ministry of Transport.



Credit: By U.S. Air Force photo by Edward Aspera Jr. - United States Air Force,

One of the core tasks of meteorological services is the warning of dangerous weather conditions. These include above all thunderstorms, as these are often accompanied by storms, hail and heavy rains. The German Weather Service uses the system "NowcastMIX" for this purpose. It polls several remote sensing systems and observation networks every five minutes to warn of thunderstorms, heavy rain and snow over the next two hours. 

"However, NowcastMIX can not detect thunderstorm cells until heavy precipitation already occurs. Therefore, using satellite data, one tries to detect the formation of the thunderstorm cells earlier, in order to warn accordingly earlier, "explains Professor Jens Dittrich, who teaches computer science at the Saarland University and heads the group" Big Data Analytics ". Together with his doctoral student Christian Schön and the meteorologist Richard Müller of the German Weather Service, he has therefore developed a system that could soon complement NowcastMIX in the prediction of thunderstorms. Her project is a first step in exploring the applicability of artificial intelligence in the prediction of weather and climate phenomena.

In order to be able to accurately predict thunderstorms in a certain region, the so-called convection of air masses, ie the rise of heated air with simultaneous decrease of colder air in the environment must be recognized early and precisely. This has been known for a long time. The clou of the new system, however, is that it requires only two-dimensional images, namely satellite imagery, to detect these three-dimensional air shifts.

In order to recognize on the two-dimensional images what is happening three-dimensionally in the sky, the researchers use photographs taken fifteen minutes apart. Part of the series of images for each area is input to an algorithm that calculates what the future, unentered image would look like. The scientists then compare this result with the real image. The magnitude of the difference between prognosis and reality, which the researchers call "the error", then serves as input for a second algorithm, which the researchers used to train with machine learning to detect the relationship between the size of an error and the appearance of a thunderstorm.

 In this way, they can calculate whether it flashes and thunders or not. "That's the strength when we apply artificial intelligence to large amounts of data. It recognizes patterns that are hidden from us, "explains Professor Dittrich. This is another reason why he and his colleagues have just initiated the new bachelor and master program "Data Science and Artificial Intelligence".

In lightning and thunder this combination is definitely "promising", so Dittrich. "Based on the satellite imagery alone, we can predict lightning with 96 percent accuracy for the next 15 minutes. If the prediction window is opened further, the accuracy is reduced, but it still remains above 83 percent for up to five hours. "However, the false-alarm rate is still too high, the researchers say. However, they believe that they can significantly reduce this if they train their model for other features, for example, also uses the currently used system NowcastMIX. To investigate this in more detail, the Federal Ministry of Transport has already granted the computer scientists from Saarbrücken 270,000 euros.



Contacts and sources:
Gordon Bolduan

Saarland University


Citation: The Error is the Feature: How to Forecast Lightning using a Model Prediction Error". ACM SIGKDD 2019: https://bigdata.uni-saarland.de/publications/SDM19_SIGKDD.pdf




No comments:

Post a Comment