Early warning systems are pivotal in mitigating risks from extreme weather events, such as floods. Leveraging artificial intelligence can significantly enhance forecast accuracy in under-monitored regions.

To understand the severity of this phenomenon, reflecting on the events of the past year is revealing. In Italy, following an extended drought, the northeastern regions, particularly Emilia-Romagna, suffered severe flooding. The initial heavy rains fell on arid soil, which could not absorb the water, leading to overflowing riverbanks. Subsequent torrential downpours caused 17 deaths, displaced thousands, and inflicted billions of euros in damage, devastating agricultural lands, businesses, and infrastructure.

Looking beyond national borders, the situation appears even more dire. In September of the previous year, a sequence of extreme meteorological events struck various regions globally. The month began with a typhoonbattering Hong Kong, resulting in significant damage and floods. Soon after, the Mediterranean faced Storm Daniel, a rare “medicane” (Mediterranean hurricane), with hurricane-like features, that caused destruction across Greece, Turkey, and beyond.

The most catastrophic of these storms occurred in Libya, where intense rainfall led to the collapse of two dams, resulting in over 4,000 deaths and leaving thousands missing. Concurrently, in Asiatwo typhoonswrought extensive damage in Taiwan and Hong Kong, and significant rainfall affected both Brazil and the United States.

In conclusion, floods are the most frequent type of natural disaster. The urgent need for action and preventative strategies is underscored by the fact that the incidence of flood-related disasters has more than doubled since the year 2000.

millennium, an increase attributed to the acceleration of the hydrological cycle driven by climate change. A NASA study has indicated that the proportion of people residing in areas vulnerable to flooding has significantly risen, pointing to a wider-reaching impact than was previously forecasted by climate models.
Artificial Intelligence (AI) is increasingly proving to be an essential tool in combating the effects of climate change, greatly improving both the precision and lead times of weather predictions, especially for extreme events such as floods. AI is adept at identifying areas at high risk and facilitating the development of mitigation strategies, thereby helping to lessen the effects of natural disasters.
Early warning systems for floods have been demonstrated to drastically reduce both loss of life and economic damage, particularly in countries with low to medium income levels. International research has underscored that improving the extent and speed of these systems can prevent thousands of deaths each year. The success of these systems also hinges on the ability to issue timely and intelligible alerts, a domain where AI can make substantial contributions by analyzing data in real-time and producing automated alerts.

The scenario

The rise in both the frequency and severity of these events is intricately linked to an accelerating hydrological cycle, driven by anthropogenic climate change.

This connection is underscored by a NASA study which reveals that the proportion of the global population living in flood-prone areas increased from 20% to 24% from 2000, a tenfold increase over predictions from earlier models. Climate change is exacerbating extreme rainfall, sea level rise, and hurricane intensity.

Early warning systems serve as a vital measure to mitigate flood risks. In this context, two World Health Organization studies (“The Global Climate 2001–2010: A Decade of Climate Extremes” and “The Global Climate 2011-2020: A Decade of Acceleration”) demonstrate that such systems can decrease flood-related fatalities by up to 43%.

Further studies, notably those published by the United Nations and the World Bank, have long pointed out that alarms and early warning systems can reduce the economic impact of floods by 35% to 50%.

Moreover, beyond isolated incidents, there is a broader aspect that warrants consideration: nearly 90% of the 1.8 billion people at risk of flooding reside in low- and middle-income countries. Although economic activities are largely concentrated in higher-income nations, the impact of floods is disproportionately greater in lower-income environments.

Returning to a pivotal point: the World Bank notes that upgrading flood early warning systems in developing countries to the standards of developed countries could save an average of 23,000 lives annually.

Climate change and negligence as causes of flooding

Heavy rainfall, storm surges, and glacier melt: these elements of climate change undeniably contribute to flooding, enhancing their likelihood, frequency, and intensity. Yet, there is more to consider.

Land management also plays a critical role, for instance by removing natural features that would typically slow and facilitate the absorption of rainwater. Road paving, for example, impedes water absorption, while deforestation increases water runoff and landslide risks.

Significantly, the study “Late-stage Deforestation Enhances Storm Trends in Coastal West Africa” specifically investigated deforestation in the South-Western African region, ongoing since 1900 over a coastal stretch of 300 km.

Utilizing three decades of satellite data, it underscores how deforestation has intensified convective activity. An increase in the frequency of afternoon storms has been observed both above and downstream from deforested areas, with more substantial increases in larger zones.

Near the coast, where sea breeze convection prevails, storm frequency in deforested areas has doubled due to the increased thermal contrast between land and sea. This is particularly significant in rapidly growing cities such as Freetown and Monrovia, which face heightened risks of sudden flooding. Ongoing deforestation along the coast could thus continue to amplify storm activity.

The role of AI in early warning systems

In this context, artificial intelligence offers tangible support through data analysis. This concept isn’t new, though recent developments in research have brought innovative advances.

AI is proving to be an indispensable tool in combating the effects of climate change, particularly with regard to extreme weather events. It plays a vital role in identifying high-risk areas and in developing strategies for mitigation, resilience, and response.

Crucially, AI underpins the development of early warning systems that alert communities to imminent disasters, providing invaluable time for preparation and evacuation.

AI and Machine Learning are also pivotal in managing the aftermath of catastrophic events, for instance, by employing imaging systems to compare aerial photographs before and after an event. This helps in determining where interventions are needed, assessing damage, and coordinating relief and reconstruction efforts.

A case in point is the deep learning model DAHiTrA, which classifies building damage using satellite imagery collected after natural disasters. This model recognizes the geographical features of different locations and compares pre- and post-disaster images of buildings, roads, or bridges to assess damage levels. This aids authorities in quickly determining the number of affected buildings and infrastructures and the extent of the damages the day following the event.

Machine Learning models are also employed to analyze data from social media, which provide insights that help report issues such as interruptions in water or electricity supplies, thus supporting emergency management operations.

Forecasts for unmonitored basins

As previously mentioned, researchers are now turning their attention to entirely new areas.

Generally, it is important to recognize that hydrological forecasting models require calibration for each river basin using extensive historical data. Challenges arise when dealing with basins lacking hydrometers, making it impossible to gather data.

This significant problem has been the focus of the International Association of Hydrological Sciences (IAHS) for a decade through the PUB project – “Prediction in ungauged basins”. After ten years of research, the IAHS found that progress was limited and noted that the most significant advancements had occurred «in monitored basins, not unmonitored ones, which adversely affects developing countries the most». Recent studies start from a stark reality: only a small fraction of the world’s river basins are monitored, and hydrometers are not uniformly distributed globally.

Early warning systems: a project for predicting rare events

According to the recent study “Global Prediction in Ungauged Watersheds,” conducted by a collaboration of American and European researchers, there exists a robust correlation between a country’s Gross Domestic Product (GDP) and its public accessibility to river flow observation data. This correlation indicates that high-quality analysis and forecasting in regions vulnerable to flooding are profoundly more challenging.

The study leveraged machine learning technologies to develop adaptable hydrological simulation models for basins without monitoring, aiming to establish a global forecasting system utilizing publicly available data. The goal was to generate precise, large-scale river flow predictions, with a particular focus on extreme events.

The latest technology for real-time global hydrological forecasts is epitomized by version 4 of the Global Flood Awareness System (GloFAS). Managed by the Copernicus Emergency Management Service (CEMS), GloFAS is under the stewardship of the European Commission’s Joint Research Centre (JRC) and operated by the European Centre for Medium-Range Weather Forecasts (ECMWF).

Various forecasting systems exist worldwide, and many nations have dedicated agencies responsible for early warning dissemination. Considering the devastating impact of floods globally, it is imperative for forecasting agencies to assess and benchmark their forecasts, alerts, and methodologies. A pivotal initial step in this direction is the archiving of historical forecasts.

The Artificial Intelligence (AI) model developed in this study uses Long Short-Term Memory (LSTM) networks to forecast daily river flow across a seven-day forecast horizon. The model underwent training and testing on external samples employing random k-fold cross-validation at 5,680 flow measurement stations [note: K-fold cross-validation is a validation method that evaluates the performance of a statistical model by partitioning the initial dataset into \( k \) equal-sized segments, or folds. Each iteration uses a different fold as the test set while the remaining \( k-1 \) folds serve as the training set. This cycle is repeated \( k \) times, with each of the \( k \) folds being used exactly once as the test set. The aggregate of evaluations from each iteration gives a robust estimate of the model’s effectiveness].

Contrastingly, GloFAS metrics were computed using a combination of monitored and unmonitored sites and across various calibration and validation periods, thus enabling comparisons with the GloFAS benchmark. This methodology was necessitated by the high computational costs of recalibrating GloFAS, precluding the feasibility of restarting calibration with each new validation set.

The objective is to ascertain the reliability of extreme event forecasts; hence, in this particular study, the concept of “return periods” is critically significant. Return periods denote the expected frequency of extreme events, often more difficult to predict using traditional hydrological models. Concerns persist that the reliability of data-driven models may wane when facing rare events poorly represented in training datasets.

Nevertheless, the comparative analysis of the GloFAS system and AI-based forecasts demonstrates that AI maintains superior accuracy and reliability, predicting events with return periods ranging from one to five years up to five days in advance.

This underscores AI’s capability to adeptly manage the forecasting of crucial hydrological events and its potential to significantly enhance natural disaster prediction, thus becoming an invaluable asset for extreme event mitigation and response strategies.

Beyond generating accurate forecasts, the prompt dissemination of alerts is also crucial. Consequently, all analyses and reprocesses utilized for this study have been made available in an open-source repository, alongside a research iteration of the machine learning model within the NeuralHydrology repository on GitHub.

There remains significant scope for enhancement in global flood prediction and early warning systems, which are essential for the safety and well-being of millions worldwide. These improvements could dramatically affect the lives and properties of those at risk from floods. In light of this, an open invitation is extended to researchers and organizations with access to river flow data to contribute to the open-source Caravan project.

Glimpses of Futures

We aim to broaden the scope of these discussions, focusing on anticipating potential future scenarios and detailing, via the STEPS matrix, the varied impacts that advancements in early warning systems may have.

S – SOCIAL: the UNDRR’s Sendai Framework for Disaster Risk Reduction 2015-2030 sets out seven targets and four action priorities. Target 7 specifically aims to “substantially increase the availability and access to multi-hazard early warning systems and disaster risk information and assessments for people.” The UNDRR (United Nations Office for Disaster Risk Reduction) maintains that an effective, people-centered, end-to-end early warning system comprises four interlinked components: (1) disaster risk knowledge based on systematic data collection and assessments; (2) hazard monitoring, analysis, and forecasting; (3) the dissemination and communication of timely, accurate, and actionable warnings and related information about likelihood and impact from an official source; (4) preparedness at all levels to act on the received warnings. Increasingly, studies underscore the importance of social preparedness in mitigating losses from floods, particularly where technical forecasting and alert capabilities are limited.

T – TECHNOLOGICAL: the study highlighted here shows that using artificial intelligence and open data sets can greatly improve the accuracy and lead time of short-term (0-7 days) forecasts for extreme river events. On average, the reliability of current immediate global forecasts has been extended to a lead time of five days. Notably, by implementing an AI-based methodology, it has been possible to achieve forecast accuracy in regions with sparse monitoring, such as Africa, at levels comparable to those in Europe.

E – ECONOMICResearch by the JRC a few years ago illuminated the diverse direct and indirect impacts of natural disasters like floods, which span multiple time scales. During such events, human lives and property are endangered, with civil protection measures primarily aimed at safeguarding people, homes, and essential infrastructure in the short term. Damage to transport networks, including ports and airports, as well as to energy infrastructure, can lead to medium- to long-term disruptions adversely affecting the competitiveness of local, regional, and national industries. Additionally, floods can overwhelm sewer systems and wastewater treatment facilities, leading to potential releases of untreated water into the environment, contaminating drinking water supplies and fostering conditions favorable for the spread of waterborne diseases. Moreover, floods facilitate the release of toxic substances such as gasoline, pesticides, detergents, and paints, causing enduring environmental contamination. The resulting damage to productive activities not only escalates unemployment rates but also fosters a growth in investor mistrust, triggering a cascading economic downturn in the impacted region. In a globalized world, the repercussions of major disasters are likely to extend beyond local and national confines, potentially disrupting international industries, infrastructures, and supply chains with lasting effects.

P – POLITICAL: approximately one-third of the world’s population still lacks early warning systems, particularly in less developed countries and developing island states. Against this backdrop, a significant commitment has been made by the United Nations. In 2022, UN Secretary-General António Guterres declared the UN’s pledge to ensure that by 2027, every person on Earth will be protected by early warning systems. UNESCO, a pioneer in developing early warning systems for various natural hazards, particularly tsunamis, plays a vital role in this initiative and works with member states to achieve this goal. Over recent decades, UNESCO has expanded its early warning capabilities to include additional risk types such as floods, droughts, wildfires, and glacier melts. Enhancing the availability of multi-hazard early warning systems and disaster risk information remains one of the seven global objectives outlined by the previously mentioned Sendai Framework for Disaster Risk Reduction (2015-2030).

S – SUSTAINABILITY: climate risk early warning systems must be founded on strong scientific and technical bases, focusing on the individuals or sectors most at risk. This requires adopting a systematic approach that encompasses all pertinent risk factors—those arising from climatic hazards as well as social vulnerabilities—and considers both short-term and long-term processes. Globally, active early warning systems facilitate monitoring, prediction, and alerting of populations about various hazards including tropical cyclones, floods, storms, tsunamis, avalanches, tornadoes, cloudbursts, volcanic eruptions, extreme temperature events, wildfires, and droughts. Plans for early warning and disaster relief are crucial for mitigating the impacts of extreme events. However, ongoing climate change and poor land management are likely to increase the frequency and intensity of these phenomena, necessitating more effective actions on this front.

Written by:

Maria Teresa Della Mura

Giornalista Read articles Look at the Linkedin profile