Integrating AI and climate change scenarios for multi-risk assessment in the coastal municipalities of the Veneto region

Integrating AI and climate change scenarios for multi-risk assessment in the coastal municipalities of the Veneto region Maria Katherina Dal Barco a,b , Margherita Maraschini a,b , Ngoc Diep Nguyen a,b , Davide Mauro Ferrario a,b , Olinda Rufo a,b , Heloisa Labella Fonseca a,b , Sebastiano Vascon a,c , Silvia Torresan a,b,* , Andrea Critto a,b a Department of Environmental Sciences, Informatics and Statistics, Ca’ Foscari University of Venice, Venice, Italy b CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy c European Centre for Living Technology (ECLT), Ca’ Foscari University of Venice, I-30123 Venice, Italy HIGHLIGHTS GRAPHICAL ABSTRACT •Determine the daily risk score and estimate the annual frequency of impacts. •Linear regression enables the estimation of annual risks in coastal areas. •Sea-level is responsible for rising trend, precipitation for seasonal variability. •Gradual increase in expected annual impact trend across all scenarios. ARTICLE INFO Editor: Dami` a Barcel´ o Keywords: MLP Linear regression Climate change Coast Impacts Future scenario ML ABSTRACT Global climate is experiencing exceptional warming, leading to a rise in extreme events worldwide. Coastal regions are particularly vulnerable to climate change (CC), due to dense populations, interconnected economies, and fragile ecosystems. These areas face escalating risks as CC intensifies the severity and frequency of extreme weather phenomena, like heavy precipitation, sea-level rise (SLR), storm surges. Integrated approaches are crucial to assess the combined impacts of atmospheric and marine hazards at the land-sea interface. Machine Learning (ML) offer innovative solutions to analyse multi-risk events, leveraging large and heterogeneous datasets and modelling complex, non-linear interactions. This study introduces a two-tier ML approach to estimate risks associated with extreme weather events for the Veneto coastal municipalities under current and future scenarios. The model, tested and validated with presentday data, showed satisfactory performance (error margin ~20 %). The model was applied to mid-term (until 2045) and long-term (until 2100) periods under different CC scenarios, represented by various Representative * Corresponding author at: CMCC Foundation - Euro-Mediterranean Center on Climate Change, Italy. E-mail addresses: [email protected] (M.K. Dal Barco), [email protected] (M. Maraschini), [email protected] (N.D. Nguyen), [email protected] (D.M. Ferrario), [email protected] (O. Rufo), [email protected] (H.L. Fonseca), [email protected] (S. Vascon), [email protected] (S. Torresan), [email protected] (A. Critto). Contents lists available at ScienceDirect Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv https://doi.org/10.1016/j.scitotenv.2025.178586 Received 16 October 2024; Received in revised form 17 January 2025; Accepted 18 January 2025 Science of the Total Environment 965 (2025) 178586 Available online 30 January 2025 0048-9697/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/bync-nd/4.0/ ). Concentration Pathways (RCP). Mid-term analysis reveals an increasing risk trend, driven by SLR under RCP8.5, underscoring the significance of considering non-linear interactions between multiple marine and atmospheric hazards. Long-term analysis highlights how future risks depend mainly on precipitation and SLR across the analysed CC scenarios (RCP2.6/4.5/8.5). Results indicate a gradual increase in the expected annual risk trend, with RCP8.5 scenario showing the most severe outcomes. By 2100, the risks under RCP8.5 are projected to be ten times higher than those observed during the historical period, highlighting the importance of developing effective strategies to address these challenges. 1. Introduction Over the past three decades, global temperatures have soared to unprecedented levels, driving a significant increase in frequency and intensity of extreme events worldwide (IPCC, 2023). Coastal regions are particularly vulnerable to these climate-change-induced impacts due to the combination of high population density, interconnected economic activities, fragile ecosystems, and the complex interactions of multiple hazards occurring across varying temporal and spatial scales. Among these, rising sea surface temperatures, a direct consequence of climate change, have played a pivotal role in amplifying the severity of these extreme events. In the Atlantic, hurricanes have become increasingly frequent and intense, with devastating consequences for coastal communities (M´ endez-Tejeda and Hern´ andez-Ayala, 2023; Moharana and Swain, 2023). Hurricane Katrina in 2005 highlighted the catastrophic socioeconomic and environmental consequences of such events (Diaz et al., 2020; Raker and Woods, 2023). It caused massive economic losses and became one of the deadliest disasters in recent history, underscoring the far-reaching consequences of climate-induced hazards (Diaz et al., 2020; Shek-Noble, 2023). Similar patterns are emerging in Europe, where the increasing occurrence of extreme marine and weather events is a growing concern (Weilnhammer et al., 2021). For instance, in 2018, Storm Adrian caused extensive damages across the Alpine region before reaching the Adriatic Sea a couple of days later, demonstrating the transboundary nature of such events (Biolchi et al., 2019; Menegatto et al., 2024). More recently, the Storm Dana in autumn 2024 caused extraordinary devastation in the Valencian coastal community, resulting in severe damage, significant loss of life, and exposing critical gaps in disaster management systems. 1 These events, driven by both marine and atmospheric hazards, marked a turning point highlighting the urgent need for a multi-hazard approach to achieve effective risk assessment and management (Hochrainer-Stigler et al., 2023; Ward et al., 2022). This approach is essential to address the escalating threats that vulnerable coastal regions are facing. Accordingly, an integrated approach is crucial to assess the interplay among all interacting risk factors (i.e., hazard, exposure, and vulnerability), as well as to evaluate the impacts that may occur at the land-sea interface. Recently, Artificial Intelligence (AI), particularly Machine Learning (ML) algorithms, has gained significant traction, offering a new path to tackle the analysis of these multi-risk events. These algorithms excel at integrating great volumes of heterogeneous data and modelling non-linear relations between multiple factors. The growing potential of data-driven methodologies is poised to transform climate risk assessment (Kashinath et al., 2021; Reichstein et al., 2019; Zennaro et al., 2021). By integrating diverse sources of information, ML approaches provide an unprecedented opportunity to capture complex multi-hazard interactions and evaluate their broader impacts on social, economic, and natural systems (Reichstein et al., 2019). For example, ML models have been widely used in studies analysing environmental-health risks, such as the combined effects of heat and air quality stressors (Boudreault et al., 2023; Guo et al., 2024; Wang et al., 2019). Similarly, population displacement has been studied through ML algorithms that investigate the dynamic relationships between droughts, conflicts, and food security (T´ arraga et al., 2024). Additionally, significant progress has been made in developing innovative models to generate climate simulations at different spatio-temporal scales (Gualdi et al., 2013), and to assess the physical impacts of climate change (Park and Lee, 2020; Pham et al., 2023; Saha et al., 2021; Sarhadi et al., 2017). To the best of authors knowledge, impacts have been modelled at global, regional and local scales worldwide (Tabari and Willems, 2023; Wang and Yan, 2021; Park and Lee, 2020). However, there remains a significant gap in multi-risk assessment at the local scale for the Veneto coast. While individual coastal risks, such as coastal erosion and water quality degradation, have been extensively studied and analysed (Dal Barco et al., 2024; Fogarin et al., 2023; Pham et al., 2023, 2024), no comprehensive studies have been conducted to address the combined risks in this region. This gap is particularly evident in the North Adriatic region, characterised by complex interactions among various hazardous events (e.g., sea-level rise, extreme precipitations, storms). Despite Dal Barco et al. (2024) analysed risks in the coastal municipalities of the Veneto region during the 2009–2020 timeframe, no studies have yet estimated the cumulative impacts of climate change starting from daily values of hazard indicators, or projected future trends in impact occurrence. Building upon the gaps, this paper introduces a novel two-tier ML approach designed to estimate the annual occurrence of impacts in the Veneto coastal area, and the subsequent application of the developed algorithm to future climate change scenarios. The aim of the methodology is the calculation of risk as the number of days in which at least an impact caused by extreme weather events occurs in the Veneto coastal area at municipality level. Although the application of binary classification algorithms for the identification of rare events is common to several subjects areas (Malik and Ozturk, 2020), such as weather predictions (de Coste et al., 2022; Rao et al., 2020), social science (Hain and Jurowetzki, 2020), political science (Hegelich, 2016), finance (Coffinet and Kien, 2019), manufacturing failure (Dogan and Birant, 2021), fraud detection (Yousefi et al., 2019), and medical examinations (Luca et al., 2014), the following step - i.e., estimation of the cumulative number of these events - is never discussed. The calculation of the cumulative value is a challenging task due to the skewness of the input dataset (i.e., the small number of days with impacts over the total number of days) that characterises rare events identification problems (Liu and Bondell, 2019). In the frame of this application, building upon the work discussed in Dal Barco et al. (2024), a multilayer perceptron (MLP) algorithm with weighted cross-entropy loss has been implemented to determine the daily risk score for each municipality, hence quantifying the likelihood of an impact to occur at the given location and day. This daily risk score is then used to estimate the annual frequency of days on which at least one impact is expected along the coastal municipalities. The algorithms were then implemented, trained, validated and tested on reanalysis data. The algorithm was then applied to climate change scenarios, in order to evaluate the contribution of different hazards, the non-linearity effects, and to estimate how climate change will influence the impact frequency in the Veneto coastal area in the years to come. 1 Spain floods: Death toll rises to 205 as nation braces for more rain – Kieran Guilbert for Euronews (31/10/2024): https://www.euronews.com/2024/10/3 1/several-missing-in-spain-after-heavy-rain-causes-flooding M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 2 2. Description of the case study The study area covers the eleven coastal municipalities of the Veneto region (Fig. 1B), facing the North Adriatic Sea along a coastline of about 169 km (Ruol et al., 2016). These municipalities are located in the provinces of Venice (i.e., the Metropolitan city of Venice) and Rovigo (ISTAT, 2021), with the Po and Tagliamento rivers marking the southern and northern borders of the region, respectively. The Veneto region features a highly heterogeneous landscape (Fig. 1A), resulting in distinct hazards impacting various areas of the region (Zonn et al., 2021). The mountainous zones (related to the provinces of Belluno, Verona, and Vicenza) are predominantly affected by landslides, particularly during heavy rainfall events (Franceschini et al., 2022; Meena et al., 2022). The inland plains, included in the provinces of Padova, Verona and Vicenza, are primarily affected by river floods, which are exacerbated by intense precipitation and snowmelt (Mel et al., 2020; Pantalona et al., 2021). The water quality degradation is of critical concern in the Lagoon of Venice, due to anthropogenic pressures and hydrological modifications (Aslan et al., 2022; Zennaro et al., 2023). Finally, the coastal area of the Veneto region, the focus of this study, faces multiple hazards, including pluvial and coastal floods, sea-level rise, coastal erosion, and storm surges (Carbognin et al., 2010; Cavaleri et al., 2020; Fogarin et al., 2023; Lionello et al., 2021; Pham et al., 2024; Regione del Veneto, 2024). The coastal region is divided into three distinct morphological zones (Bezzi et al., 2018). The northern part is characterised by straight, sandy, and low-lying littorals shaped by the Sile, Piave, Livenza, and Tagliamento rivers (Zonn et al., 2021). The central area is dominated by the Venetian Lagoon whose original morphology was altered by historical river diversions. The Sile and Brenta rivers, which once flowed into the lagoon, were redirected during the Middle Ages to mitigate frequent catastrophic floods (Bellizia et al., 2023). Today, the lagoon is delimited by the sandy barrier islands of Lido di Venezia and Pellestrina (Zonn et al., 2021). The southernmost part of the Veneto coast features the Po river and its Delta, a complex system of river branches, outlets, and salt marshes, recognised as one of the most ecologically significant areas in the Veneto region (Ruol et al., 2018; Torresan et al., 2008). Despite multiple sediment inflow, the Veneto coastline has been undergoing massive coastal erosion due to urbanisation and anthropic pressure during the last century, considerably changing the overall natural land-use due for urbanisation, industrialisation, and tourism purposes. To the present day, a total of about 60 km of Veneto coastline preserves its natural condition, as a result of the lagoon and river estuary areas that characterise it (Legambiente, 2012). Specifically, the littoral includes natural protected areas, regional parks, and reserves, which are embedded in the European ecological network Natura 2000, such as the Delta Po regional park, among others (Regione Regione Veneto, 2012; Ruol et al., 2016). However, the improper management of the areas carried out during the last century has decreased the sedimentation budget, which was exacerbated by climate change related hazards (e.g., sea-level rise, storm surge). In this perspective, the regular implementation of adaptation measures (e.g., beach nourishment, dune restoration) is supporting the recovery of coastal areas from both natural and anthropogenic pressures (MATTM, 2017). From a socio-economic point of view, the regional capital is driven by fisheries, aquaculture, agriculture, industrial activities, maritime traffic, offshore activities and tourism (Torresan et al., 2012). The latter Fig. 1. Case study area: (A) the geographical context of the Veneto region, showing its location in northeastern Italy, and (B) the coastal municipalities of the Veneto region analysed in this study. M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 3 is made possible by the mild climate that the Veneto region offers throughout the year due to its proximity to the Adriatic Sea, as well as the presence of diversified landscapes (e.g., mountains, lakes, hills) and numerous UNESCO sites across the territory (Cuccia et al., 2016). Looking at the coastal area of the Veneto region, it belongs to the subcontinental temperate zone (Barbi et al., 2013), with higher average annual temperatures (about 14 ◦C) than the inland areas (around 13 ◦C; Barbi et al., 2013). Moreover, the coastal area experiences fewer rainy days and lower rainfall accumulation compared to the rest of the region (Barbi et al., 2012). However, it records a higher frequency of days with heavy precipitation, with average annual rainfall ranging between 700 and 1000 mm (Barbi et al., 2012). The historical exposure of the Veneto coastal area to various natural and anthropic pressures has been exacerbated even more in recent decades by climate change. In particular, the rising mean temperature (about 0.57 ◦C per decade across the Veneto region; ARPAV, 2023) and the increasing frequency of heatwave events, combined with the intensification of heavy rainfall, strong winds, and extreme sea-level events, are expected to exacerbate health risks for the population (Ferrarin et al., 2013, 2022; Umgiesser et al., 2021). Under this already complicated scenario, the North Adriatic basin, and especially the Veneto coastal area, is expected to experience an increasing trend of extreme atmospheric (e.g., heatwaves, storms) and marine events (e.g., sea-level rise, storm surges, inundations) in the forthcoming decades (Lionello et al., 2020; Pham et al., 2023; Zanchettin et al., 2020). 3. Data collection The development of the proposed two-tier ML model requires the collection of the daily information on impacts and extreme events that occurred across the coastal municipalities of the Veneto region during the period 2009–2019 (Section 3.1). Additionally, estimates of future extreme events indicators, will be derived from two sets of future climate change models: atmospheric and marine projections (Sections 3.2.2 and 3.2.1, respectively). The indicators providing insights into potential impacts were adapted from Dal Barco et al. (2024) and includes daily maximum sea surface height, the daily cumulative precipitation, the maximum value of the daily cumulative precipitation in the previous 3 days, the maximum daily precipitation in a month and maximum daily wind velocity. on the data for historical daily indicators and extreme events (2009–2019) will form the reference dataset. Indicators derived from future climate change models will be classified between baseline dataset, for indicators relative to the historical period, and future dataset, estimated for dates after 2019. To create the baseline and future datasets, two types of scenario models will be employed: high-resolution models incorporating all hazard indicators for mid-term analysis (until 2045), four models based on varying climate change scenarios for longterm analysis (until 2100). Data source for each indicator is summarised in Table 1. 3.1. Historical data collection of hazards and impacts for the creation of the reference dataset The reference dataset compiles information on atmospheric and marine hazards, as well as recorded impacts, with daily temporal resolution and municipality-level spatial resolution. Atmospheric input data were collected from the Copernicus European Regional ReAnalysis (CERRA and CERRA Land; Schimanke et al., 2021; Verrelle et al., 2022). Compared to global reanalysis products, the CERRA dataset offers a higher horizontal resolution, enabling a more local-scale applications and a more accurate representation of topography and physiographic features (Verrelle et al., 2022). Marine hazard data were obtained from the Copernicus Marine Service (CMEMS), part of the European Union’s Earth Observation Programme. In particular, hourly sea surface height data were retrieved from the Mediterranean Sea Physics Reanalysis database (MEDSEA_MULTIYEAR_PHY_006_004) 4 , and validated using local data from the Acqua Alta platform. The maximum sea surface height indicator was calculated with respect to the tide-level zero of Punta della Salute, Venice. Finally, the impact dataset was created by collecting qualitative information from the ‘State of crisis’ (Stato di crisi) reports, which document damages and services interruptions caused by extreme events in the coastal municipalities of the Veneto region. This dataset comprises pairs of (day, municipality) where at least one impact was reported, excluding economic damage data due to its unavailability in public records. However, the dataset presented several limitations. Reports covered the entire Veneto region and listed affected municipalities and event dates without always specifying the exact day each municipality was impacted. Extreme events often caused inland impacts days before reaching the coast, leading to inaccuracies in the recorded impact days. To improve accuracy, each recorded impact day was manually crosschecked with local newspaper reports to resolve discrepancies. The final dataset includes 447 days of recorded impacts caused by Table 1 Summary of data sources for machine learning application in the Veneto region. Indicator Timeframe Spatial domain Spatial resolution Temporal resolution Source Precipitation indicators [mm] 2009–2019 Europe 5.5 km ×5.5 km daily CERRA 1 2009–2045 North-East Italy 6 km ×6 km daily AdriaClim-WRF 2 2009–2100 North-East Italy 5 km ×5 km daily EURO-CORDEX 3 Maximum wind velocity [m/s] 2009–2019 Europe 5.5 km ×5.5 km 3-h CERRA 1 2009–2045 North-East Italy 5 km ×5 km daily AdriaClim-WRF 2 2009–2100 – – – – Maximum sea surface height [m] 2009–2019 Mediterranean Sea 1/24◦(~4 km) hourly, daily CMEMS 4 2009–2045 Mediterranean Sea 1/24◦(~4 km) 6-h CMEMS 4 +NASA 5 2009–2100 Mediterranean Sea 1/24◦(~4 km) 6-h CMEMS 4 +NASA 5 Impacts 2009–2019 Veneto region Municipality daily Veneto Region 6 1 Copernicus European Regional ReAnalysis (CERRA): Schimanke et al., 2021; Verrelle et al., 2022. 2 Model developed in the frame of the Interreg Italy-Croatia AdriaClim (Climate change information, monitoring and management tools for adaptation strategies in Adriatic coastal areas) project: https://programming14-20.italy-croatia.eu/web/adriaclim. 3 Models provided by EURO-CORDEX (Coordinated Downscaling Experiment - European Domain): https://euro-cordex.net/. 4 Mediterranean Sea Physics Reanalysis provided by Copernicus Marine Service (CMEMS), and developed at the Euro-Mediterranean Center on Climate Change (CMCC): https://doi.org/10.25423/CMCC/MEDSEA_MULTIYEAR_PHY_006_004_E3R1. 5 Sea-level rise rate estimated by National Aeronautics and Space Administration (NASA): Fox-Kemper et al., 2021; Garner et al., 2021, 2023. 6 Veneto Region – Post-Emergency Disaster Events Management Office: https://bur.regione.veneto.it/BurvServices/Pubblica/sommarioDecretiPGR.aspx?expand=19 M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 4 extreme weather events between 2009 and 2019, covering the eleven coastal municipalities of the Veneto region. Despite these efforts, detailed typologies of damages could not be associated with specific municipalities and were thus excluded from the analysis. 3.2. Modelled climate indicators for the creation of the baseline and future datasets Estimating future impacts require projections of climatic hazard indicators, provided by marine and atmospheric models. Projections for the historical period will form the baseline dataset, which will be used to validate the methodology and assess the quality of the results. Finally, projections for future periods will be employed to estimate future impacts. Ideally, a single model capable of simulating both marine and atmospheric components and accounting for their cascading interactions would be used. However, no suitable unified model was available for this case study focusing on coastal municipalities of the Veneto region. As a result, separate datasets were selected for atmospheric and marine indicators. The baseline and future datasets for marine indicators were created using sea-level trends calculated by NASA and distributions derived from reference data. For atmospheric indicators, different models were selected based on their suitability for midand long-term analyses. The AdriaClim-WRF model, developed under the Interreg Italy-Croatia AdriaClim project (Fedele et al., 2024), was used for the mid-term risk analysis (up to 2045), whereas EURO-CORDEX simulations were used for the long-term estimation of impacts (up to 2100). The choice of models reflects their distinct advantages. AdriaClim-WRF model provides detailed information on daily precipitation rates and maximum wind velocity, which is crucial for understanding hazard interactions that lead to impacts. On the other hand, EURO-CORDEX models, while lacking wind velocity data, cover a longer timeframe and offer projections for three different climate scenarios. The estimation of indicator projections considered various climate change scenarios, such as Shared Socioeconomic Pathway (SSP) and Representative Concentration Pathways (RCP) scenarios. According to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC), RCP scenarios are denoted as ‘RCPy’, where ‘y’ represents the approximate radiative forcing level (in watts per square metre, Wm–2) by the year 2100 ″ , and SSP scenarios are referred as ‘SSPx-y’, where ‘x’ specifies to socio-economic trends, and ‘y’ indicates the radiative forcing level in 2100 (IPCC, 2023; van Vuuren et al., 2011). Table 2 outlines the scenarios used for the midand long-term projections, while the methodologies behind the atmospheric and marine models is detailed in the following Sections 3.2.1 and 3.2.2, respectively. For consistency, the application uses ‘RCPy’ notations to identify the scenarios. 3.2.1. Atmospheric models 3.2.1.1. AdriaClim-WRF model. The AdriaClim-WRF model was developed under the AdriaClim project to estimate mid-term coastal risks (up to 2045) under the RCP8.5 climate change scenario. This model utilises climate downscaling from regional (MED-CORDEX) to subregional (AdriaClim-WRF) scales, driven by one of the coupled air-sea models over the Mediterranean region provided by the Med-CORDEX coordinated initiative (Ruti et al., 2016), specifically the LMDZNEMOMED model (L’H´ ev´ eder et al., 2013). AdriaClim-WRF model provides a comprehensive set of atmospheric variables, including temperature, wind, precipitation, humidity, covering the 1990–2045 timeframe (Fedele et al., 2024). 3.2.1.2. EURO-CORDEX models. For the long-term risk analysis (up to 2100), atmospheric indicator projections were selected from available EURO-CORDEX simulations. These projections align with the IPCC’s RCP scenarios, reflecting different greenhouse gas emission trajectories and socio-economic trends (IPCC, 2023). Climate change studies integrate these scenarios to represent multi-sectorial indicators (e.g., emissions of greenhouse gases and air pollutants, energy and land use, socioeconomic and technological changes, among others; Vuuren et al., 2011). In this study, RCP2.6 (low emissions), RCP4.5 (medium emissions) and RCP8.5 (high emissions) scenarios were chosen, to represent a broad range of emission pathways. The simulations, with an original spatial resolution of 0.10◦ (approximately 11 km), provide daily data on precipitation, mean and maximum temperatures, and atmospheric pressure for the 1970–2100 period. Only precipitation was selected for the analysis of future scenario analyses. Each EURO-CORDEX simulation comprises a regional climate model (RCM) with a global climate model (GCM) and includes different emission pathways (Giorgi, 2019), as listed in Table 3. 3.2.2. Marine model Projections of the maximum sea surface height indicator for both midand long-term applications were derived by combining sea-level trends provided by NASA with distributions from reference data. This method assumes no significant impact of climate change on the variability of sea-level values around the average, consistent with current practices (Buchanan et al., 2016, 2017; Raicich, 2018; Rasmussen et al., 2018; Tebaldi et al., 2012, 2021; Wahl et al., 2017). The procedure to create maximum sea surface height projections for each municipality includes the following steps: - Calculate the average trend of reference maximum sea surface height values. - Determine the residuals as the difference between each day’s maximum sea surface height value and trend value. - Estimate future sea-level trends for each scenario using NASA’s sealevel increase rate (Fox-Kemper et al., 2021; Garner et al., 2021, 2023). - Consider three scenarios based on SSP and RCP combinations: SSP1–2.6, SSP2–4.5, and SSP5–8.5. Table 2 Summary of the Shared Socioeconomic Pathway (SSP) and Representative Concentration Pathways (RCP) scenarios selected for the estimation of indicator projections in the midand long-term periods. Total daily precipitation Maximum wind velocity Maximum sea surface height Midterm RCP8.5 RCP8.5 SSP5–8.5 Longterm RCP2.6 –SSP1–2.6 RCP4.5 –SSP2–4.5 RCP8.5 –SSP5–8.5 Table 3 List of EURO-CORDEX simulations selected for the long-term risk analysis. ‘Acronym’ column includes the names of the simulations as referred hereinafter. Acronym Global model Regional climate model RCA4 EC-EARTH (Irish Center for High-End Computing, Ireland) RCA4 (Swedish Meteorological and Hydrological Institute, Sweden) RACMO22E RACMO22E (KNMI, Netherlands) CCLM CCLM4–8-17 (CLM com, EU) REMO MPI-ESM-LR (Max Planck Institute for Meteorology, Germany) REMO2009 (Max Planck Institute in System and Control Theory, Italy) M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 5 - Randomly extract a residual value for each day in the future period, ensuring alignment with the corresponding municipality, calendar month and lunar phase from the reference dataset. - Combine scenario-specific sea-level trends with selected residual to generate the baseline dataset (2009–2019) and future datasets (2020−2100) for each municipality and scenario. 4. Proposed methodology 4.1. Machine Learning model development Machine Learning (ML) algorithms offer a new path to address the analysis of multiple environmental hazards and to model future climate change impacts due to their ability to model complex feedback and nonlinear interactions between different variables, without the need for an explicit modelling. In Dal Barco et al. (2024), the authors analysed the risk of impacts along the coastal municipalities of the Veneto region within the 2009–2020 timeframe and developed a ML tool able to estimate, for each day and municipality, a risk score representing the likelihood of an impact. However, the analysis does not include any estimation of future impact trends. This application starts from the methodology described in Dal Barco et al. (2024) and aims to develop a two-tier ML methodology capable of estimating the number of days per year in which at least an impact occurs along the Veneto coastal area given the knowledge of daily local atmospheric and marine indicators or indicator projections. This tool will provide an insight of the effects of climate change on impact trends. The input data of the analysis is composed of a set of environmental features for each day in the timeframe 2009–2019 and for each municipality on the Veneto coastal area, and by a boolean variable representing the presence of an impact on the same day and location that will be used as output label (detailed in Section 3.1). As the purpose of this application is the estimate of the number of days per year with impacts, the dataset should be structured with yearly samples. In this way, the input dataset would be composed of 10 samples (one per year), and with more than 20,000 features (i.e., 11 municipalities ×number of indicators ×365 days), and could not be used to train any supervised ML tool, as the number of input features would be much bigger than the number samples. To overcome this problem, several different approaches have been considered, including time series, use of statistics of indicators, convolutional neural networks. Eventually, a novel two-tier approach was developed (Fig. 2). In the first tier the number of features in each yearly sample is strongly reduced by estimating with a first ML tool a risk score value for each couple (day, municipality); this step is a binary classification performed using a supervised ML algorithm and it is described in detail in Dal Barco et al. (2024). The results of the first tiers are then manipulated to create the input of the second tier: the daily risk score data are grouped in yearly samples using a moving window and binned in risk classes. Then, in the second tier, the number of days per year in each class are the input of the linear regression tool which calculates the total number of days featured by the occurrence of impacts associated to extreme weather events per year. This method leveraged well-known ML algorithms to exploit the realtime relationship between indicators and impacts effectively. The novelty of the proposed approach does not rely on the implemented ML methods, but on the way these aggregating methods are combined, which allowed the integration of daily input features and impacts to calculate the cumulative yearly values of impacts despite the skewness of the input dataset. 4.1.1. Tier1: ML algorithm to evaluate the daily risk score The first tier aims at estimating the risk score for each day and municipality, which is a measure of the probability of an impact occurring at that location on that day. Since this first step follows the previous application of the authors (Dal Barco et al., 2024), the methodological approach is not discussed in detail. Each sample of the input dataset is associated with a day and a municipality: the input features are the atmospheric and marine indicators (Table 1), plus the municipalities and the months of the year, and the output label is the presence/absence of impacts. This is a binary classification problem on a skewed dataset (as the number of days with impact is about 1 % of the total number of days). A Multi-Layer Perceptron algorithm has been trained with sample weights to keep the minority class in the right account, validated, and tested. The output of the application on a new input sample is the corresponding risk score, which is a measure of the probability of the occurrence of an impact on that (day, municipality). It is worthwhile to note that the outcome of this algorithm is as reliable as its input data: when it comes to look at future climate projections, the results, as the corresponding modelled data used as input, will not reliable on a daily basis, however their statistical distribution over several years, and hence the number of days with high values of risk score, is reliable. Because of the skewness of the input data and the consequently use of weights in the codes, it is not possible to calculate the number of days with impacts per year as the sum of the risk scores. Similarly, it is not possible to set a threshold in the risk score, define any value with risk score above the threshold as day with impact and sum them up: in a binary classification on a skewed data if the recall is sufficiently high (i. e., most positive events are correctly identified), the precision is likely to be very low (i.e., there will be a high number of false positive). This makes the total amount of predicted positive an overestimate of the correct number of positives. These considerations led to the creation of a second algorithm, that takes as input the daily risk scores and output the number of days with impact per year. Fig. 2. Scheme of the two-tier approach. M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 6 4.1.2. Tier2: ML algorithm to estimate annual extreme weather impact The second tier is aimed at estimating the number of days per year in which at least one impact is expected along the Veneto coastal area using as input the knowledge of the daily risk scores previously calculated. In order to achieve this goal, the dataset was structured using yearly samples. The feature of each sample is characterised by the 365 values of the risk score for the 11 municipalities, i.e., approximately 4000 features; the output label is the number of days with at least an impact on the Veneto coastal area. The total number of samples is 10: the number of full years in the timeframe. As the number of features is much smaller than the number of samples, to create a ML algorithm it is necessary to reduce the number of features and increase the number of samples. The increase of the number of samples was obtained by the implementation of moving windows sampling: instead of considering 10 nonoverlapping years, a year was defined as any interval of 12 consecutive months. Since samples adjacent in time are not independent from each other, making a random splitting a non-viable option, the data was split chronologically between train and test before the moving window process. To reduce the number of features, the risk scores of each sample were binned with a user-defined number of classes (i.e., risk classes), and the number of samples falling in each bin was calculated. The number of events per year falling in each class are the features of the yearly samples. Finally, a linear regression algorithm with non-negative coefficients and zero-intercept was implemented estimating the number of annual impacts derived from the number of days in each risk class. The algorithm is then trained, validated and tested, and the results are reported in Section 5.2.1. 4.2. ML application to future scenario data for the estimate of future impacts In order to implement the developed ML methodology to the estimate of future impact, the data must meet the requirements for a correct application (section 4.2.1). Then, the ML model is applied to mid-term modelled hazard data to analyse the non-linear effects between hazards (section 4.2.2), and finally to multiple climate change models until year 2100 (section 4.2.3). Fig. 3. Distribution of (a) maximum daily wind velocity, (b) total daily precipitation, and (c) maximum daily precipitation in a month from the CERRA reanalysis against AdriaClim-WRF and EURO-CORDEX datasets (under different climate change scenarios – i.e., RCP2.6, RCP4.5, RCP8.5) within the historical period (2009–2019). M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 7 4.2.1. Comparison of the distributions of the reference data and the modelled data in the reference timeframe The consistency of the dataset for reference and baseline in the reference timeframe (2009–2019) needs to be checked. In particular, the data used to train, validate, and test the ML algorithms should have a distribution as close as possible to the distribution of the modelled data used to estimate future impacts. In other words, the algorithm can be applied to new data (i.e., the modelled data) as long as the distributions of the extreme values of the reanalysis and the modelled data are spatially and temporally comparable. The distribution between reference and baseline values of maximum sea surface height are the same because of the methodology used to create the modelled data, which is based on a random sampling of the reference values (see section 3.2.2). More attention must be devoted to the atmospheric indicators: for the reference datasets these indicators were retrieved from the CERRA dataset, while the modelled data are retrieved by the AdriaClim-WRF model for the mid-term predictions and for the EURO-CORDEX models for the long-term predictions (see Section 3.2.1). Fig. 3 shows the distributions of both reference and baseline datasets for three selected atmospheric indicators within the reference timeframe 2009–2019. For the wind indicator, higher values are observed in the baseline dataset compared to the reference (Fig. 3-a). For both rain indicators, instead, the reference data present a thicker tail distribution, highlighting a slightly higher precipitation level (Fig. 3-c, Fig. 3-d). However, the distributions of the two datasets are comparable, which allow the further application of ML algorithms to modelled data. 4.2.2. Mid-term risk analysis with synergistic interaction of hazards using the AdriaClim-WRF model The main aim of this application is to understand the contribution of the different marine and atmospheric indicators (i.e., precipitation, sealevel, wind velocity) in relation to the annual probability occurrence of impacts for the upcoming decades. This analysis will provide information on the relative importance of each potential hazard and their combination (Teichert et al., 2016). The comparison of coastal impact trends across different combinations of hazards provides a more comprehensive understanding of how climate change will exacerbate these impacts. This analysis explores the influence of individual hazards, their interactions, and the effects of the non-linearities in the system. Fig. 4 shows the combination of hazards identified for this study, and Section 5.3 presents the results of future annual risk estimations. The input of these analyses are the indicator projections for daily maximum sea surface height, the daily cumulative precipitation, the maximum value of the daily cumulative precipitation in the previous 3 days, the maximum daily precipitation in a month and maximum daily wind velocity, which are considered the main responsible for coastal impacts along the Veneto littoral, especially within the 2009–2019 timeframe (Dal Barco et al., 2024). When a hazard is accounted, the related indicator projections assume the values calculated by climatic models following the RCP8.5 climate change scenarios. Conversely, when a hazard is not considered, the indicator projections are set to the average of the indicator values in the reference period. In other words, if a hazard is excluded, the corresponding indicator assumes realistic and non-extremes conditions, ensuring that the hazard cannot contribute to any impacts predicted by the model. The outputs of these analyses are the future trends of coastal impacts until year 2045 for each hazard combination; their comparisons highlight the relative importance of the stressor and uncover non-linearity effects of compounding hazards. 4.2.3. Long-term risk analysis for multiple climate change models using the EURO-CORDEX models In order to estimate future impacts dependent on climate change scenarios, hazard indicators modelled for different Representative Concentration Pathway (RCP) scenarios were required to be integrated into the developed ML algorithm. In particular, the scenarios considered for this application were RCP 2.6, RCP4.5 and RCP 8.5. For each of the RCP scenario, one maximum sea surface height dataset was created, as described in Section 3.2.2. As far as the atmospheric hazard indicators are concerned, the four models introduced in Section 3.2.1 were considered. In particular, for each model, datasets from the three RCP scenarios were extracted. Combining both marine and atmospheric datasets, a total of 12 input datasets were created. Both tiers of the algorithm were then applied to the 12 datasets, and the corresponding impact trends for the Veneto coastal area were estimated. On top of the four models, for each scenario the results for an ensemble model were calculated. In order to create the ensemble model for each scenario, the option of averaging the indicator data for each day of the timeframe was quicky discarded, as its smoothing effect would have reduced the extreme values, which are the main drivers of the Fig. 4. Iterative combination of the selected hazards to estimate future risks. M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 8 impacts. Conversely, the ensemble model for each scenario was created by considering the results of Tier 1 of the developed ML model for each of the four models and calculating the number of days following in each class. Then, the number of days in each of the high-risk class was averaged between models, and the average number of days per class was used as input the second algorithm. 5. Application of the methodology to the Veneto coastal area As described in Section 4, the results of the application of the proposed two-tier ML-based approach to the Veneto coastal area are presented in the following sections. In particular, the model validation and testing for the estimation of the annual occurrence of impacts is presented in Section 5.1, whereas Section 5.2 is devoted to the application of the two-tier approach with modelled data over the reference period, followed by its application to different climate change scenarios estimating the occurrence of impact risks along the coastal municipalities of the Veneto region (Italy) in the midand long-term future (Sections 5.3 and 5.4, respectively). 5.1. ML model testing The training and the testing of the proposed algorithms are performed separately for the midand the long-term analysis, as the latter does not include the wind indicator, and the algorithm needs to be trained and tested again without it. Fig. 5 depicts the training and testing of the algorithm developed for the mid-term analysis, comparing the recorded (in blue) and predicted (in orange) impact frequency over time, whereas Fig. 6 illustrates the same results for the long-term analysis. The results obtained by the algorithms for both analyses are quite similar. Only a slight performance loss for the results of the long-term analysis is noted, as the wind indicator information is missing. There is a good agreement between both sets (i.e., between the number of predicted and recorded impacts): the similar quality of the results between train and test excludes the risk of overfitting. The least square error (i.e., the median of the squared distance between the number of predicted and recorded impacts) on the test sets is approximately 2 impacts for both midand long-term analysis, which is approximately the 20 % of the average number of impacts per year. The determination coefficient R 2 relative to the correlation between recorder and predicted number of impacts on the test set is 0.57 for the long-term analysis and 0.60 for the mid-term analysis. This result is quite satisfactory, given the uncertainty of the input data, hence the tested algorithm is now ready to be applied to new data. 5.2. Application of the two-tier approach to estimate the annual occurrence of impacts Once the Machine Learning (ML) models are implemented and tested in the reference period, they can be applied to new input data. In order to estimate future occurrence of impacts in relation to climate change scenarios, modelled data were integrated, only after having checked that their distribution in the reference period is similar to the one of data used for training and testing (Section 4.2.1). 5.2.1. Validation of the impact predictions In order to validate the results of the algorithm, the ML models are applied to the baseline datasets - i.e., datasets containing projection of the indicator values evaluated during the historical timeframe, when the recorded impacts are available. ML models have been applied to the baseline datasets created by both AdriaClim-WRF and EURO-CORDEX baseline datasets (described in Section 3.2), and their results have been compared with the recorded impacts (Fig. 6 and Fig. 7). It can be noted that the outputs of the three different climate change scenarios (i.e., RCP2.6, RCP4.5 and RCP8.5) do not significantly differ during the historical period (Fig. 7). This is expected, as the projected indicators for the three scenarios are almost identical in the historical timeframe, diverging gradually with time in the future timeframe. The results of these analysis are affected by two distinct sources of errors: the errors introduced by the ML algorithms (discussed in Section 5.1), and the error induced by modelling of input data. The error due to the ML algorithm (i.e., distance between blue and orange dots) is much smaller than the total error (i.e., distance between green and orange dots), meaning that the main approximation introduced by the application of the tool is related to the modelling of the input data, rather than to the algorithm. More specifically, both figures show how the results obtained by modelled data are not able to predict the yearly variability of the collected data, and in general underestimate or misplace the number of days with impacts. The analysis of the indicators’ distribution (Fig. 3) highlights that modelled indicators present lower precipitation levels compared to reanalysis. This is likely to be the cause of the underestimation of impact,. However, the order of magnitude of the yearly impact is comparable. These results suggest that the application of the algorithm to future modelled data allows a reliable estimation of the trends of the impacts, with a possible underestimation or misplacement of the peaks caused by precipitation extremes. 5.3. Mid-term risk analysis with synergistic interaction of hazards Once the algorithm for mid-term prediction was tested on the reference period, it has been applied to the AdriaClim-WRF model for the 2020–2045 timeframe to estimate future risks. To better understand the contribution of the different hazards in the calculation of the annual Fig. 5. Comparison of recorded number of impacts, number of impacts predicted by the mid-term algorithm, number of impacts predicted by the longterm algorithm: (A) train dataset; (B) test dataset. M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 9 2016. Med-CORDEX initiative for Mediterranean climate studies. Bull. Am. Meteorol. Soc. 97 (7), 1187–1208. https://doi.org/10.1175/BAMS-D-14-00176.1. Saha, A., Pal, S.C., Santosh, M., Janizadeh, S., Chowdhuri, I., Norouzi, A., Roy, P., Chakrabortty, R., 2021. Modelling multi-hazard threats to cultural heritage sites and environmental sustainability: the present and future scenarios. J. Clean. Prod. 320, 128713. https://doi.org/10.1016/j.jclepro.2021.128713. Sarhadi, A., Burn, D.H., Yang, G., Ghodsi, A., 2017. Advances in projection of climate change impacts using supervised nonlinear dimensionality reduction techniques. Climate Dynam. 48, 1329–1351. https://doi.org/10.1007/s00382-016-3145-0. Schimanke, S., Ridal, M., Le Moigne, P., Berggren, L., Und´ en, P., Randriamampianina, R., Andrea, U., Bazile, E., Bertelsen, A., Brousseau, P., Dahlgren, P., Edvinsson, L., El Said, A., Glinton, M., Hopsch, S., Isaksson, L., Mladek, R., Olsson, E., Verrelle, A., Wang, Z.Q., 2021. CERRA sub-daily regional reanalysis data for Europe on single levels from 1984 to present. Copernicus climate change service (C3S) climate data store (CDS). https://doi.org/10.24381/cds.622a565a. Shek-Noble, L., 2023. Communicating vulnerability in disasters: Media coverage of people with disabilities in hurricane Katrina and the T¯ ohoku earthquake and tsunami. In: The Palgrave Handbook of Disability and Communication. Springer International Publishing, pp. 295–316. https://doi.org/10.1007/978-3-031-144479_18. Tabari, H., Willems, P., 2023. Global risk assessment of compound hot-dry events in the context of future climate change and socioeconomic factors. Npj Climate and Atmospheric Science 6 (1), 74. https://doi.org/10.1038/s41612-023-00401-7. T´ arraga, J.M., Sevillano-Marco, E., Mu˜ noz-Marí, J., Piles, M., Sitokonstantinou, V., Ronco, M., Miranda, M.T., Cerd` a, J., Camps-Valls, G., 2024. Causal discovery reveals complex patterns of drought-induced displacement. IScience 27 (9), 110628. https://doi.org/10.1016/j.isci.2024.110628. Tebaldi, C., Strauss, B.H., Zervas, C.E., 2012. Modelling Sea level rise impacts on storm surges along US coasts. Environ. Res. Lett. 7 (1), 014032. https://doi.org/10.1088/ 1748-9326/7/1/014032. Tebaldi, C., Ranasinghe, R., Vousdoukas, M., Rasmussen, D.J., Vega-Westhoff, B., Kirezci, E., Kopp, R.E., Sriver, R., Mentaschi, L., 2021. Extreme Sea levels at different global warming levels. Nat. Clim. Chang. 11, 746–751. https://doi.org/10.1038/ s41558-021-01127-1. Teichert, N., Borja, A., Chust, G., Uriarte, A., Lepage, M., 2016. Restoring fish ecological quality in estuaries: implication of interactive and cumulative effects among anthropogenic stressors. Sci. Total Environ. 542 (Part A), 383–393. https://doi.org/ 10.1016/j.scitotenv.2015.10.068. Torresan, S., Critto, A., Dalla Valle, M., Harvey, N., Marcomini, A., 2008. Assessing coastal vulnerability to climate change: comparing segmentation at global and regional scales. Sustain. Sci. 3 (1), 45–65. https://doi.org/10.1007/s11625-0080045-1. Torresan, S., Critto, A., Rizzi, J., Marcomini, A., Mendez, F.J., Leschka, S., FraileJurado, P., 2012. Assessment of coastal vulnerability to climate change hazards at the regional scale: the case study of the North Adriatic Sea. Natural Hazards & Earth System Sciences 12 (7). Umgiesser, G., Bajo, M., Ferrarin, C., Cucco, A., Lionello, P., Zanchettin, D., Papa, A., Tosoni, A., Ferla, M., Coraci, E., Morucci, S., Crosato, F., Bonometto, A., Valentini, A., Orli´ c, M., Haigh, I.D., Nielsen, J.W., Bertin, X., Fortunato, A.B., Nicholls, R.J., 2021. The prediction of floods in Venice: methods, models and uncertainty (review article). Natural Hazards & Earth System Sciences 21 (8), 2679–2704. https://doi.org/10.5194/nhess-21-2679-2021. van Vuuren, D.P., Edmonds, J., Kainuma, M., Riahi, K., Thomson, A., Hibbard, K., Hurtt, G.C., Kram, T., Krey, V., Lamarque, J.-F., Masui, T., Meinshausen, M., Nakicenovic, N., Smith, S.J., Rose, S.K., 2011. The representative concentration pathways: an overview. Clim. Change 109 (5). https://doi.org/10.1007/s10584011-0148-z. Verrelle, A., Glinton, M., Bazile, E., Le Moigne, P., Randriamampianina, R., Ridal, M., Berggren, L., Und´ en, P., Schimanke, S., Mladek, R., Soci, C., 2022. CERRA-land subdaily regional reanalysis data for Europe from 1984 to present. Copernicus climate change service (C3S) climate data store (CDS). https://doi.org/10.24381/cds. a7f3cd0b. Wahl, T., Haigh, I.D., Nicholls, R.J., Arns, A., Dangendorf, S., Hinkel, J., Slangen, A.B.A., 2017. Understanding extreme sea levels for broad-scale coastal impact and adaptation analysis. Nat. Commun. 8, 16075. https://doi.org/10.1038/ ncomms16075. Wang, J., Yan, Z., 2021. Rapid rises in the magnitude and risk of extreme regional heat wave events in China. Weather and Climate Extremes 34, 100379. https://doi.org/ 10.1016/j.wace.2021.100379. Wang, Y., Song, Q., Du, Y., Wang, J., Zhou, J., Du, Z., Li, T., 2019. A random forest model to predict heatstroke occurrence for heatwave in China. Sci. Total Environ. 650, 3048–3053. https://doi.org/10.1016/j.scitotenv.2018.09.369. Ward, P.J., Daniell, J.E., Duncan, M., Dunne, A., Hananel, C., Hochrainer-Stigler, S., Tijssen, A., Torresan, S., Ciurean, R., Gill, J.C., Sillmann, J., Couasnon, A., Koks, E., Padr´ on-Fumero, N., Tatman, S., Lund, M.T., Adesiyun, A., Aerts, J.C.J.H., Alabaster, A., de Ruiter, M.C., 2022. Invited perspectives: a research agenda towards disaster risk management pathways in multi-(hazard-)risk assessment. Nat. Hazards Earth Syst. Sci. 22, 1487–1497. https://doi.org/10.5194/nhess-22-1487-2022. Weilnhammer, V., Schmid, J., Mittermeier, I., Schreiber, F., Jiang, L., Pastuhovic, V., Herr, C., Heinze, S., 2021. Extreme weather events in europe and their health consequences – a systematic review. Int. J. Hyg. Environ. Health 233, 113688. https://doi.org/10.1016/j.ijheh.2021.113688. Yousefi, N., Alaghband, M., Garibay, I., 2019. A Comprehensive Survey on Machine Learning Techniques and User Authentication Approaches for Credit Card Fraud Detection. https://doi.org/10.48550/arXiv.1912.02629. Zanchettin, D., Bruni, S., Raicich, F., Lionello, P., Adloff, F., Androsov, A., Antonioli, F., Artale, V., Carminati, E., Ferrarin, C., Fofonova, V., Nicholls, R., Rubinetti, S., Rubino, A., Sannino, G., Spada, G., Thi´ eblemont, R., Tsimplis, M., Umgiesser, G., Zerbini, S., 2020. Review Article: Sea-Level Rise in Venice: Historic and Future Trends. Natural Hazards and Earth System Sciences Discussions, November, pp. 1–56. https://doi.org/10.5194/nhess-2020-351. Zennaro, F., Furlan, E., Simeoni, C., Torresan, S., Aslan, S., Critto, A., Marcomini, A., 2021. Exploring machine learning potential for climate change risk assessment. Earth Sci. Rev. 220, 103752. https://doi.org/10.1016/J.EARSCIREV.2021.103752. Zennaro, F., Furlan, E., Melaku Canu, D., Aslan, S., Solidoro, C., Critto, A., 2023. Evaluation of lagoon eutrophication potential under climate change conditions: A novel water quality machine learning and biogeochemical-based framework, 24–28 Apr 2023 2023. EGU General Assembly, pp. EGU23–8409. https://doi.org/10.5194/ egusphere-egu23-8409. Zonn, I.S., Kostianoy, A.G., Semenov, A.V., Joksimovi´ c, A., Đurovi´ c, M., 2021. Veneto, pp. 363–364. https://doi.org/10.1007/978-3-030-50032-0_604. M.K. Dal Barco et al. Science of the Total Environment 965 (2025) 178586 16