Through the new technology, which was developed by researchers at Saarland University in Germany, the scientists hope to be able to predict storms with greater accuracy than ever before.
Lightning is a common occurrence in most parts of the world. Often pretty to watch, the risk of injury can be very high. In the case of the researchers’ own country, June 2019 saw 177,000 lightning bolts in the night sky occurring over just a few days. The outcome was that many people were injured, either directly by the lightning or by the associated storm.
In order to improve the ability of meteorologists to predict problematic weather, the researchers are constructing a system to better predict local thunderstorms. The approach uses satellite images together with artificial intelligence, with the aim of combining image analysis with physical data in order to train the artificial intelligence and hence develop an accurate model.
Currently the Deutscher Wetterdienst, who assess weather across Germany, make use of the “NowcastMIX” system. The technology takes a sample every five minutes, drawing data from several remote sensing systems and observation networks. An attempt is made to assess likely storms over a two-hour window. The weakness is that the predictive power only works when heavy precipitation has already started.
The new system, by making use of digital satellite images, will be able to issue warnings much further in advance. The images enable an assessment by studying the convection of air masses (the rise of heated air while colder air sinks in the surrounding are). The artificial intelligence is able to take two-dimensional images and process these to detect these three-dimensional air shifts.
To make the predictions robust, the first image series for a given area is inputted to an algorithm, which calculates what the future image would look like. The computer scientists then compare this result with the real image. The extent of the deviation between the forecast and reality serves as input for a second algorithm. The second algorithm is trained using machine learning to help assess the relationship between the size of the error and the occurrence of a thunderstorm. This process gradually reveals more and more patterns, and helps to improve the predictive capabilities of the model.
The work to date has been promising, and the researchers have been granted 270,000 additional euros, from the Federal Ministry of Transport, in order to fund the next phase of the study.
