Large amounts of data and new methods and technologies with which to analyze them. The new frontier of machine learning, a branch of artificial intelligence, at the service of climate studies, in research by the CMCC Foundation and Ca’ Foscari University of Venice

Global warming is exacerbating weather and climate extreme events. The interaction between different forms of hazards triggered by climate change will cause future cross-sectoral impacts affecting a variety of natural and human systems.

Research can improve the understanding of these interactions and dynamics, support decision-makers in managing current and future climate change risks, also thanks to an improved ability to predict expected risks and quantify their impacts.

To this end, in recent years, the scientific community has started testing new methodological approaches, technologies, and tools, among which the application of machine learning, which can help exploit the potential of large amounts and variety of environmental monitoring data available today (big data).

What are the results of the exponential increase in the application of machine learning methods for the assessment of climate-induced risks?

In the study “Exploring machine learning potential for climate change risk assessment“, a team of scientists from the CMCC Foundation and Ca’ Foscari University of Venice conducted an in-depth review of more than 1,200 articles on the subject, published in the last 20 years, highlighting the potential and limitations of machine learning in this field.

“Machine learning is a branch of artificial intelligence,” explains Federica Zennaro, a researcher at the CMCC Foundation and Ca’ Foscari University Venice and the main author of the study. “By simulating the processes of the human brain, certain mathematical algorithms can understand the relationships between a set of input data in order to predict the required output. In our research, we identified that floods and landslides are the most analyzed events through machine learning models, probably because they are the most relevant and frequent around the world.”

Moreover, the study reveals that machine learning has two major potentials that make it particularly interesting when applied to this field of study.

The first is that said algorithms can learn from data: the more data, the better algorithms learn. Thanks to its ability to analyze and process large amounts of data, machine learning allows researchers to disentangle complex relationships underlying the functioning of socio-ecological systems, exploiting the big data collected from various sources, including sensors for environmental analysis at high temporal frequency, social media, satellite data and images, and drones.

The second is that they can combine different types of data, thus enabling an assessment of the risk extent whilst taking into account all its dimensions. These include not only the triggering hazard (for example, an increase in rainfall) but also the vulnerability and exposure of the socio-economic system at stake, which are crucial factors in an evaluation of overall impacts

“For example, consider a model that is trained with detailed data on flood events over the past 20 years, including their location and information on the affected context (urban or natural). This model can project, in a scenario characterized by future climate conditions, what the probability of an event happening at a certain point will be, and calculate its risk of causing harmful impacts to society and the environment,” Zennaro explains. “Machine learning represents the future of risk assessment, but its great potential is not yet widely exploited. Our research shows that there are still few studies that use these models to develop long-term future risk scenarios (up to 2100). The vast majority of studies focus on the short term, probably influenced by the reduced availability of extended time series data capable of supporting adequate model training for long-term projections.”

The next step, explains the co-author Elisa Furlan, a researcher at the CMCC Foundation and Ca’Foscari University Venice, is to develop machine learning models that are increasingly efficient at studying and untangling the complex spatiotemporal interrelationships among different climatic, environmental, and socioeconomic variables, thereby improving understanding of the behavior of complex systems. “Under the perspective of a rising abundance of data and machine learning models’ complexity, researchers will have the possibility (and duty) to improve the understanding of climate-related risks, with the main aim of providing accurate and sound multi-risk scenarios able to drive robust adaptation planning and disaster risk reduction and management”.