AI studies nature

 By Sara Sangermani

In the past few weeks, nature has shown all its power in several episodes. The eruption of Stromboli, the powerful earthquake in California, the violent rainfall, hailstorms and tornadoes that have affected many Italian locations, raise the question: was it possible to predict these phenomena?

Being able to anticipate potentially dangerous natural phenomena means saving human lives, limiting damage and being able to act immediately to combat the emergency. The numbers speak for themselves: worldwide, in 2018 alone, the various natural disasters caused  a total damage of 160 billion dollars. The highest price was paid mainly by the United States, because of the hurricane Florence that last September hit and devastated the eastern part of the country.
Trying to predict these natural phenomena seems therefore the best solution to try to limit as much as possible the damage. Experts from the various sectors have been working for some time to understand how to anticipate the various typologies of natural phenomena. So many are looking at the potential of Artificial Intelligence (AI) but the events to be monitored are not all the same and there are still limits and problems to overcome.

Predict volcanic activity

The sudden awakening of Stromboli caught everyone by surprise: could one foresee that? According to the experts, the notice could only be given a few minutes in advance. The limited availability of data on the activities of some volcanoes, including Stromboli, means that it is not possible to create predictive models to anticipate a destructive event. However, there are some parameters that are constantly monitored and that allow us to understand a few minutes before if an eruptive activity is about to begin. Little time, perhaps only enough to block any excursions at the start but not to save people already close to the crater. Then there is the problem of the impossibility of calculating the extent of an eruption, such as the moment when the lava spills or the fall of pyroclastic material.
The volcanoes, however, are not all the same and there are some, such as Etna, whose continuous activity has allowed experts over the years to accumulate many data. Also, Etna is a volcano where magma is in contact with the atmosphere hence, is possible to record sounds one hour before the beginning of the eruption. A recent study started in 2010, has shown that for volcanoes similar to the Sicilian one it is better not to rely on AI but to analyze the variations of low frequency acoustic waves.   The results of this study are more than positive: out of 59 events monitored, there were only two false alarms. It is a precious tool both for not setting off excursions and for the safety of air traffic.

Night-time image of the erupting Etna

Predict seismic activity

Another catastrophic event that is difficult to control is the seismic activities of the subsoil. Also in this case, in fact, the major problem is represented by the small amount of data collected, which makes it difficult to build and adopt predictive models. Despite this, AI can be particularly useful in the post-earthquake phase. A deep learning system connected to images recorded via satellite, for example, allows in a short time to have an accurate estimate of the damage, knowing quickly the amount of buildings collapsed and the number of blocked roads. This makes it possible to start a well-organised rescue machine immediately, capable of intervening immediately and effectively.
The same principle is applied in Japan for the protection of infrastructure. By analysing satellite images acquired over different periods of time, the IA is able to detect any anomalies that could indicate a problem in an infrastructure, such as a bridge.

Predict extreme weather activities

Floods, floods, whirlwinds and hurricanes of unprecedented intensity are increasingly affecting countries that were previously unconnected with these phenomena. In this case, the use of AI is giving positive results. The data available are in fact numerous and even if you cannot predict the extent of the disaster, you know more and more with greater certainty where it will affect and so you have several hours in advance to alert the population.
NASA, for example, is working on the development of a system that allows you to track every hour the progress of a hurricane so you know exactly where it will hit. This was not possible until a few years ago, when the population was only notified every 6 hours.
Even a major player in the computer world, Google, is particularly sensitive to this problem and is using many resources for the development of an algorithm capable of predicting floods, though so far, only reliable in 75% of cases. A search that, if it were to lead to more accurate results, would help to save as many as 10 thousand lives each year and prevent the destruction of goods such as houses and cars. In addition to saving lives, in fact, these predictive systems would allow to safeguard many assets, helping the economy of countries affected by these phenomena.

After the 2017 hurricane season, the U.S. government is experimenting with the possibility of anticipating its arrival (NASA)

Limits to tackle

There are obviously some limits to be overcome when it comes to systems for the prevention of natural disasters based on AI. First of all, some of the phenomena we have just seen are influenced by the rapid climate change that is taking place. This means that while it is possible to predict the flooding of a river, thanks to the current data available, it is not possible to have an exact predictive model if these data should be modified by the rise of global warming. Another factor to consider is that these systems are not always reliable. In fact, it is man who provides the IA with the data to be processed and in doing so may not have taken into account certain variables or made mistakes. In this, machines are similar to humans: they too can make mistakes in the interpretation of a certain datum. The whole ethical part would then come into play about who would then be responsible, between man and machine, if a disastrous phenomenon should not be correctly predicted or in the case of a false alarm.
Finally, the unreliability of the system means that false alarms could be created with the more than negative consequence that they could lose their effectiveness over time.

Image cover by: J.D. Griggs, United States Geological Survey

READ MORE: Scientific breakthroughs on the launch pad by Nicholas Newman

about the author
Sara Sangermani