Artificial intelligence: when risk science meets the future

The increasingly urgent environmental and climate challenges demand innovative solutions. CIMA Research Foundation is addressing this need by integrating artificial intelligence into its research efforts for forecasting and managing environmental risks. From wildfire protection to drought monitoring, AI offers new tools - and raises new questions - for tackling global issues.

In recent years, artificial intelligence (AI) has emerged as a strategic ally for research, amplifying the ability to analyze, forecast, and manage environmental risks. CIMA Research Foundation is exploring this technology by integrating it into its research projects, particularly within doctoral studies conducted in collaboration with the University of Genoa.

"With AI and machine learning, we can analyze large volumes of data quickly, optimize processes, and create more accurate predictive models," explains Mirko D'Andrea, a researcher at CIMA Research Foundation. Indeed, artificial intelligence has the potential to push scientific research beyond traditional boundaries, overcoming the limitations of manual analyses and enabling deeper and more timely understanding of environmental phenomena. We have previously discussed Large Language Models in risk reduction; now, let's see how combining advanced mathematical models with machine learning opens new possibilities in managing climate risks, such as wildfire protection, drought monitoring, and water resource management, according to our researchers.

Algorithms serving wildfire prevention

The combination of machine learning and satellite data provides a concrete example of how AI can be applied to territorial protection. In the context of wildfire risk monitoring, CIMA Research Foundation has developed advanced predictive models that analyze risk factors, map high-risk areas with greater precision, and support preventive management of such events. Niccolò Perello, a Ph.D. researcher at the Foundation, explains that AI can account for numerous variables and their complex interactions: "The climate of a region influences its susceptibility to wildfires, meaning the propensity of the area to be affected by such phenomena," he says. "Average conditions and extremes of climatic and meteorological variables-such as relative humidity, temperature, and precipitation-directly affect the likelihood of fire ignition and spread."

However, climate alone does not determine wildfire risk. "Meteorological and climatic conditions are not the only factors at play: vegetation type and continuity, topography, and human activity are additional variables, with interactions that are often nonlinear," continues Perello. "Machine learning allows us to identify these hidden relationships and project them onto territories that may present potential vulnerabilities to wildfires, learning from past events in other locations."

One of the algorithms used by CIMA Research Foundation is Random Forest, a machine learning technique that analyzes relationships between various environmental variables to predict wildfire risk evolution. This tool also enables daily risk monitoring. Another advancement is the ML4RISICO model, which, through a variant of the Time Series Forest Classifier algorithm, monitors key parameters daily, such as the number of fires, burned areas, and COAU (Unified Aerial Operations Center) activities. Perello elaborates on how the challenges encountered during the model's training were addressed: "Wildfires are rare events, and often the available data does not cover all possible scenarios favorable for their ignition. To address this, we balanced the training data between fire cases and non-fire situations, improving the model's ability to predict these events."

AI not only allows for wildfire prediction but also enables corrective actions and real-time monitoring of territorial conditions. Perello concludes, "The ML4RISICO model optimizes daily monitoring by providing crucial data for the timely implementation of prevention and response strategies."

Advanced models for snow water resource management: AI in hydrology

In the field of hydrology, CIMA Research Foundation is pushing the boundaries of science with advanced models like SHANN, which leverage AI to monitor snow levels and their melting. Giulia Blandini, a Ph.D. researcher at the Foundation, explains how combining hydrological models with machine learning algorithms, such as LSTM neural networks, is revolutionizing water basin management.

The SHANN algorithm can analyze snow data in real time and predict water flow resulting from snowmelt. "The algorithm is trained using data assimilation techniques and then replaced during modeling. Compared to traditional techniques, it is faster, with minimal performance loss, but it opens the door to implementing expensive data assimilation techniques in operational fields," she explains.

Additionally, using Random Forest to optimize snow data has allowed the Foundation to refine the automatic cleaning process for snow depth measurements, distinguishing with high precision between snow, vegetation, and random errors. "Tools like these enable us to implement automatic quality assurance and quality control (QA/QC) strategies, improving data accuracy and, consequently, the precision of hydrological models," Blandini concludes.

Managing post-event impacts: the Recovery Gap Index

Artificial intelligence is driving a true revolution in managing extreme events and their impacts, reshaping how we address crises. A significant example of this transformation is the Recovery Gap Index (RGI), a tool developed by CIMA Research Foundation to measure the response and recovery capacities of 169 countries in the face of disasters. Leveraging advanced natural language processing (NLP) techniques, the RGI automates the analysis of over 500 documents, categorizing key information related to post-disaster recovery phases. "The use of artificial intelligence in decision-making processes is revolutionizing risk management, making it more strategic and data-driven," explains Alessandro Borre, a PhD researcher at CIMA Research Foundation. By adopting NLP, the RGI can swiftly extract patterns and key themes from large volumes of unstructured information, such as reports and articles, enabling a more precise and in-depth understanding of recovery dynamics.

This innovation not only offers a comprehensive analysis of recovery capacities but also supports the planning of more targeted policies, aiding in strengthening the resilience of affected communities. "AI also allows real-time event monitoring, providing immediate updates and recommendations," Borre continues, highlighting how artificial intelligence optimizes risk management and adapts existing tools like the RGI to specific contexts. Through this approach, risk management becomes not only more efficient but also more resilient, evolving from a reactive model to a proactive one where every decision is supported by detailed and dynamic analyses.

However, despite its numerous advantages, applying the Recovery Gap Index still presents challenges. "The primary limitation of the RGI lies in its global and national scale, which makes it less suitable for applications in specific local case studies," the researcher observes. Overcoming this limitation requires calibrating the RGI for the area of interest, tailoring it to local characteristics. "Additionally, the scarcity of detailed data, such as available resources or specific risk management information, limits its ability to accurately represent response and recovery processes."

A case study highlighting the RGI's potential and challenges is the analysis of the 2015 Nepal earthquake conducted in collaboration with the International Institute for Applied Systems Analysis (IIASA). "The RGI provided an initial assessment of response and recovery capacities, but with NLP and AI, we integrated information from various sources," Borre recounts. "This allowed us to highlight the relevance of certain subcategories, such as coordination and socioeconomic development, which proved more significant than initially indicated." The case study demonstrated the importance of calibrating the RGI to better adapt to local specificities, ensuring a more accurate assessment of post-disaster dynamics. This refinement makes the Recovery Gap Index not only a powerful analytical tool but also an indispensable resource for strategic risk management on both global and local scales.

Large Language Models (LLMs) as tools for enhancing decision support systems

Large Language Models (LLMs) represent a breakthrough in decision support, particularly in complex contexts such as emergency management. These models simplify interaction with large datasets using natural language, making it easier to access essential information quickly. As Jean-Baptiste Bove, a PhD researcher at CIMA Research Foundation and the Italian Red Cross, explains: "In emergency management, access to information is critical. However, the volume of available data is often overwhelming, making it challenging to identify the information necessary for data-driven decision-making."

Integrating LLMs into these contexts, however, comes with challenges. A significant issue is the occurrence of "hallucinations," where models produce responses that appear plausible but are, in fact, incorrect. This limitation arises because models generate responses based on patterns in their training data without necessarily verifying the accuracy of the information. "In emergency management, where decisions can have life-or-death consequences, ensuring the accuracy and reliability of LLM-generated insights is critical."

Another challenge stems from the models' design: LLMs are not specifically tailored for emergency management but for general tasks. This results in a limited capacity to interpret rapidly evolving situations or adapt to incomplete data. Furthermore, their training data is static, reducing their ability to provide updated or context-specific responses to particular events.

To address these limitations, two key techniques are being explored: Retrieval-Augmented Generation (RAG) and fine-tuning. RAG is an innovative solution that connects LLMs to external knowledge repositories, such as emergency plans or risk analysis documents. This approach ensures that model-generated responses are based on up-to-date and relevant information, avoiding errors caused by a lack of context. Fine-tuning, on the other hand, involves adapting models to specific domains by training them on curated datasets related to disaster management. "For instance," Bove explains, "fine-tuning enables models to interpret maps or geolocated data, significantly improving their ability to understand specialized terminology and context."

These advancements not only enhance the precision and reliability of LLMs but also allow for the creation of scalable and transparent tools that are essential for emergency management. "By integrating techniques like RAG and fine-tuning," Bove concludes, "we can strengthen the connection between scientific knowledge and operational decisions, revolutionizing how operators interact with complex and constantly evolving data."

Toward the horizon of AI: what lies ahead?

Despite significant progress, the path to integrating artificial intelligence into scientific research processes remains long and fraught with challenges. Key difficulties include managing big data, which requires increasingly sophisticated solutions, and adopting an interdisciplinary approach, which demands highly specialized expertise. As Mirko D'Andrea states, "Integrating AI presents various challenges, such as managing large volumes of heterogeneous data, requiring specialized skills, and addressing ethical considerations tied to these technologies." CIMA Research Foundation is tackling these obstacles with a steadfast commitment to targeted training-both technical and scientific-and developing projects that engage experts from diverse disciplines.

"AI is already transforming our approach to research. On one hand, machine learning models have enabled the development of more accurate predictive tools for managing environmental risks. On the other, the introduction of LLMs has made the process of extracting information faster and more precise, allowing for the analysis of unstructured documents and sources. Additionally, AI assists us in our daily work, from drafting project documents to writing code for our applications and research products," D'Andrea adds, emphasizing the tangible contributions of artificial intelligence across all operational aspects of scientific research.

Looking ahead, CIMA Research Foundation's goal is clear: to further enhance internal AI expertise and promote interdisciplinary collaborations, fostering synergies among various research sectors and making AI a crucial tool for addressing global environmental and climatic challenges. Advanced technologies will remain at the core of innovative, targeted solutions that not only enrich scientific capabilities but also improve the management of environmental risks, propelling scientific research to new levels of precision and effectiveness.