(Nanowerk News) Dust storms are not only a nuisance to anyone trying to keep their home clean, they pose a very real health hazard and are a major ecological problem. Respiratory problems caused by inhaling dust and other airborne particles are one of the leading causes of death worldwide. Even worse, dust particles, moving freely from country to country and continent to continent, could spread pathogens, possibly contributing to the outbreak of a pandemic. In addition, dust clouds have a very significant impact on climate: They absorb and distribute sunlight, thereby changing the Earth’s temperature, and also affect cloud properties and precipitation patterns.
Dust storms usually form in arid regions, such as the Negev, the Arabian Peninsula, the Sahara, and the deserts of North America and Asia. Winds kick small particles off the ground and, while larger sand particles sink near where a storm forms, smaller dust particles can be blown hundreds or even thousands of kilometers away.
Having early warning of dust surges can help protect vulnerable populations and prevent crop damage – and, as a bonus, save us from useless housecleaning. But the rapid development and spread of these storms, coupled with the fact that they span a large area, makes it difficult to predict when, where, and how severe they will strike.
A study recently published by Dr. Ron Sarafian, Dori Nissenbaum and Prof. Yinon Rudich of the Department of Earth and Planetary Sciences at the Weizmann Institute of Science has made a breakthrough in dust storm forecasting. Studies published in npj Climate and Atmospheric Sciences (“Multi-task deep learning for early warning of dust events applied to the Middle East”), written in collaboration with Dr. Shira Raveh-Rubin, also from the same department at Weizmann.
Initially, the researchers hoped to use the knowledge gained in the field of computer vision. Because meteorological data for dust storms could be displayed as a series of satellite images, they thought that neural networks would be able to “learn” the patterns that govern the spread of storms – just as these networks have learned to recognize videos of different animals or objects.
However, their hopes were only partially realized. A typical image consists of just the three primary colors, with a fair amount of overlap between them. The meteorological “picture”, however, consists of no less than 60 variables: data on temperature, humidity, wind speed, and so on. In addition, while computerized vision systems rely on machine learning based on archives of millions of images, there are precious few images available to artificial neural networks tasked with identifying dust storms: Israeli researchers have only 60,000 of these meteorological “films,” after gathering detailed data from satellites. and earth stations for about two decades. In this relatively limited collection, it is rare to find multiple examples of dust storms forming in the same location.
In such a case, any artificial neural network trying to learn the patterns that govern the formation of dust storms in Beersheba, for example, could experience what is known as “overfitting”. In other words, they may formulate patterns based on limited circumstances and reach incorrect conclusions when new, unstudied conditions are encountered.
To their surprise, the researchers found that forecasting could be improved by making life more difficult for artificial neural networks. They tasked the network not only with learning when dust storms were expected to reach a certain point, but also with an additional problem: tracking the much wider area where the dust was spread.
For example, to predict when a dust storm is likely to hit Beersheba, the network studies how badly the storm will affect Lebanon. Using this approach, the network has access to a much larger data set, from which it can also learn the physical and meteorological conditions in which the dust spreads.
Using data collected from all of Israel’s meteorological stations over the last 20 years, the researchers showed that during the winter and dusty spring months, they were able to forecast more than 80 percent of dust storms 24 hours in advance, and about 70 percent, 48 hours in advance. Most of the incidents that the system doesn’t predict are storms that develop rapidly in a local area, making it difficult to collect regional data that can help predict them.
Adding additional problems to the artificial neural network allows it to predict, 24 hours before their formation, more than 80% of the dust storms that develop in Israel during the dusty winter and spring.
“A network trained on data from Israel can, with some adjustments, forecast dust storms elsewhere in the Middle East and even around the world,” said Sarafian. “In addition, we have created an architecture that can help predict other rare events related to meteorological data, such as extreme rainfall or flash floods.”
Adds Rudich: “The most significant achievement of this research, which we have implemented in our follow-up studies, is the use of artificial intelligence to scan large and rich data sets and to study physical principles and atmospheric processes in a way not previously available to humans. we.”