Machine learning is a form of artificial intelligence that can learn autonomously, making new connections between the input data. It has application and is being researched in a variety of areas, including risk reduction. Mirko D’Andrea, researcher at CIMA Research Foundation, explains how
We are hearing more and more about machine learning in a variety of fields – and more and more technologies are based on it. Some are already in use: for example, spam email recognition systems make extensive use of the capabilities of this form of artificial intelligence that can learn and make new connections between the data provided. Others are being studied, as in the case of diagnostic and pharmaceutical tools. Indeed, there are a variety of fields of scientific research that can benefit and make substantial advances from machine learning, and risk management and mitigation is no exception.
How and why can machine learning be employed in our field of study, i.e. risk? What are its potentials, and what are its current limitations? We have a talk about these issues with Mirko D’Andrea, researcher at CIMA Research Foundation.
What characteristics of machine learning have made it so important and pervasive in scientific research?
The answer lies in the very function of machine learning, which is the ability of these algorithms to learn how to connect massive amounts of data and make connections between them. In this way, they enable us both to analyze large amounts of information very quickly and to find relationships among the data provided, even the least obvious ones. In addition, precisely because of this ability, machine learning also enables, in many cases, a better understanding of certain phenomena under study.
How does this apply in the field of risk reduction?
We can give an example of application based on the studies on wildfire risk prevention we have conducted in CIMA Research Foundation. For many years we have tried to create, in our models, a causal link between the characteristics of an area (topography, type of vegetation, etc.) and its susceptibility to fires. This was work that required the development of even very sophisticated algorithms. The most effective solution, however, was to rely on machine learning, which, having data regarding the region of interest and the fire history of that area, is able to identify the features that influence them the most. In this way, we were able to first create a fire hazard map for Liguria Region, and then gradually expand the area to cover the entire eastern Mediterranean region.
Are there other examples of machine learning application for risk reduction? For instance, could machine learning help with flood prediction and prevention?
CIMA Research Foundation has been undertaking these studies only recently. Anyway, generally speaking, machine learning has a wide scope of applicability in the management and reduction of various natural hazards. For example, the European Center Medium Weather Forecast (ECMWF), with which we collaborate, has recently created an infrastructure based on machine learning models for weather forecasting. In addition, there are several studies to employ machine learning in flood forecasting, based on principles similar to those we have exploited for wildfires.
Machine learning is also widely used in the field of satellite observations. In fact, many of these algorithms have their first field of application in image processing.
For example, a study is underway at CIMA Research Foundation whose goal is to integrate machine learning techniques with high-resolution Earth observation data (obtained from the Sentinel-2 satellite) to improve our understanding of drought dynamics on vegetation. The ultimate goal is to identify effective drought indicators among vegetation indices that could be derived from Sentinel-2 data for future mapping of drought status over the Italian territory.
Another use is to estimate the water content of vegetation – a necessary parameter for fire risk surveys. This data is again from Sentinel-2. However, this satellite operates in the visible spectrum: this means that we cannot analyze cloud-covered areas and have to refer to antecedent data to extract curves of possible evolution of the situation. At the modeling level, this is a very complicated process and not easily modeled by classical techniques. But we could use machine learning to analyze the history of events in that area and make a prediction about their evolution.
In the field of hydrological modeling, a study is underway to use Deep Data Assimilation techniques, which is the combination of data assimilation techniques (the process of combining observed data, such as meteorological or satellite data, with mathematical models to improve the accuracy of forecasts or simulations) with deep-learning techniques, an advanced machine learning technique that uses deep neural networks, that is, having a large number of layers and parameters. The goal is to improve the performance of hydrological models by exploiting computational capabilities of neural networks and the reduced computation time associated with these techniques.
All these are examples of numerical data analysis. Can we think about using machine learning for qualitative studies as well?
Absolutely. We have already done so as well. In fact, within the EDORA project, we produced a drought risk atlas, at the European level, under current and future climate conditions. Part of the work was based on a qualitative risk analysis: thanks to machine learning, we were able to associate certain weather-climate conditions with a historical series of impacts (such as loss of agricultural production) observed in the different socio-economic sectors analyzed. In this case, language systems such as chatGPT, come to play. These systems allowed us to extract pan-European scale drought information from newspaper articles, i.e., from datasets that do not have structured information to collect data about the phenomenon.
The use of machine learning on text analytics is another frontier field, because it allows working on large bodies of text. Part of CIMA Research Foundation’s future research will also involve the application of machine learning in the analysis of legislative frameworks. This is what we are doing with a new PhD study, which we are pursuing in collaboration with the Italian Red Cross: the aim is to analyze various unstructured datasets to turn them into information for humanitarian action.
Can we close with some thoughts on the prospects of this technology for the future? What, at the moment, are the limitations?
The biggest limitation in the use of machine learning techniques is the quality of the available data. They have to be correct and as complete, reliable and representative of the phenomenon being modeled as possible. Since machine learning relies fully on the quality of the training data, having a low-quality dataset will produce poor-quality results. Similarly, having bias in the training data will produce a system with inherent bias, which can produce erroneous outputs in certain situations.
However, CIMA Research Foundation manages a wealth of meteorological, satellite, geographic, and weather model-derived data. This wealth of quality data availability offers great opportunities for the use of these techniques.
A more general, but no less important, consideration is that we are dealing with an incredibly rapidly evolving set of technologies. Recent developments in artificial intelligence represent an inflection point, a change in the speed of technological evolution. Staying abreast of these technological advances is a major challenge in the modern world, thus also in the field of risk reduction and in the issues addressed by CIMA.