Orals: Tue, 29 Apr | Room N2
Among the several artificial intelligence global atmospheric models that have recently put forward in the literature, the ACE2 stands out as a purely data-driven stable model that can be run over millennia, providing a rather reasonable representation of ENSO and to the global greenhouse gas forcing during the recent decades. In this talk, this model is evaluated in terms of its capability to generate its own internal variability, its representation of the local probability distribution of extreme winds and precipitation, the extratropical models of variability, such as the North Atlantic Oscillation and the Pacific North-America pattern, and their teleconnections to seasonal temperature and precipitation.
Additionally, selected extratropical weather extremes have been simulated with the ACE2 model, using different initialization lead times. This set of extremes includes the Capella storm in the North Sea in January1976, the Ahr Valley flash flooding in Germany in July 2021, and others.
The ACE2 model in climate mode, trained with ERA5 reanalysis, is able to produce surprisingly realistic amplitude of internal climate variability and atmospheric teleconnection patterns involving near-surface temperature and precipitation, and atmospheric circulation in upper tropospheric levels.
In weather prediction mode, the predictability of extremes benchmarked against the ERA5 data set is limited to a lead time of 2 days, and the simulated extremes may be temporally temporally shifted by about one day. Their intensity is also somewhat weaker than the benchmark data set.
The model can also be used a extreme event attribution tool, as the model can be run under changed surface boundary conditions and greenhouse-gas forcing. A short assessment of its useful in this context will be also discussed in this presentation
How to cite: Zorita, E.: Evaluating the ACE2 model in simulating extratropical climate and weather extremes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-21246, https://doi.org/10.5194/egusphere-egu25-21246, 2025.
Heat stress represents a major risk to human health, making the development of advanced warning systems essential for safeguarding individuals and communities. Data-driven models, such as FourCastNet, PanguWeather, and GraphCast, provide rapid, accurate, and publicly accessible forecasts of meteorological variables. However, these models do not provide all the variables required to calculate thermal stress indices, such as the Universal Thermal Climate Index (UTCI). To address this limitation, this study proposes a method to estimate the UTCI for southern South America using a subset of variables available from these data-driven models. First, feature selection techniques were applied, including stepwise selection and a wrapper evolutionary approach based on the Probabilistic Coral-Reef Optimization with Substrate Layers algorithm (PCRO-SL). These techniques were used to identify key variables both at the individual grid point level and within homogeneous regions of the UTCI, defined through k-means clustering. The selected variables were then incorporated into various regression and classification models, ranging from simple linear methods to the advanced Light Gradient Boosting Machine (LGBM). The performance of these models was evaluated against the ground-truth UTCI data provided by ERA5-HEAT. Results show that the combination of PCRO-SL and LGBM yielded the most accurate UTCI estimates. Key variables identified included 2-meter temperature, specific humidity, and low-level wind components. Finally, using forecasts of these selected variables from FourCastNet, PanguWeather, and GraphCast, the method was applied to estimate the UTCI during a heatwave. Forecasts up to three days show good agreement between the observed and modeled thermal stress category. Future work will explore improvements through post-processing techniques for the meteorological variables provided by data-driven models.
Acknowledgments: This work was supported by the SAFETE project, which has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 847635 (UNA4CAREER). This work has also been partially supported by the project PID2023-150663NB-C21 of the Spanish Ministry of Science, Innovation and Universities (MICINNU), and by the EU-funded H2020 project CLINT (Grant Agreement No. 101003876).
How to cite: Collazo, S., Marina, C. M., García-Herrera, R., Barriopedro, D., and Salcedo-Sanz, S.: Forecasting heat stress using data-driven model outputs, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8699, https://doi.org/10.5194/egusphere-egu25-8699, 2025.
Anthropogenic climate change (ACC) is intensifying the frequency and severity of extreme events globally, such as extreme heatwaves and heavy precipitation. Attributing individual extreme events (EEs) to ACC is critical for assessing the risks of climate change. A common method for addressing this challenge is the pseudo-global-warming (PGW) approach, which involves removing the thermodynamic ACC signal from the initial and boundary conditions in a weather or climate model to simulate the event under preindustrial conditions. However, traditional numerical, physics-based models require substantial computational resources and expertise, often delaying attribution results. This study introduces a novel attribution method that integrates the PGW approach with cutting-edge artificial intelligence weather prediction (AIWP) models. By leveraging AIWP models, which offer rapid and efficient computations, this method significantly accelerates the process of extreme event attribution, thus copying with the demand of information on due time. The ACC signal is estimated using CMIP6 historical simulations and subtracted from the initial conditions to enable AIWP model forecasts of the event without ACC influence.
Using this hybrid approach, we quantify the impact of ACC on several recent heatwave events, including the 2018 Iberian heatwave the 2022 Pacific Northwest heatwave, and the 2023 Brazilian heatwave. Our results reveal clear ACC fingerprints in the forecasted temperature fields, showing an overall increase in the severity of these events due to climate change, but with regional differences. We further validate these findings by applying the method to a hybrid-AI atmospheric model, which quantifies the role of sea surface temperature anomalies in intensifying these extreme events.
Beyond heatwaves, this approach demonstrates its versatility by detecting ACC fingerprints in extratropical cyclones. For example, the method indicates that ACC contributed to the enhanced winds associated with the extratropical bomb cyclone Ciarán that impacted Western Europe in 2023. While the method has some limitations, such as sensitivity to initial conditions and uncertainties in CMIP6 projections, it represents a significant step forward in the rapid and accessible attribution of extreme events.
How to cite: Jiménez-Esteve, B., Barriopedro, D., Johnson, J. E., and García-Herrera, R.: Using AI-driven weather prediction models for attribution of extreme events to climate change, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4349, https://doi.org/10.5194/egusphere-egu25-4349, 2025.
Recently, deep learning (DL) techniques have been extensively applied to extreme weather prediction, which demonstrates their potential to address complex meteorological challenges. However, the success of DL-based weather prediction methods relies heavily on the availability of high-quality labelled training data. Human annotators and automated labelling tools may make mistakes due to limited expert knowledge or systematic errors, which leads to the noisy label problem. To address the noisy label challenge, we propose a novel transfer-learning-assisted cooperative sample selection (TLACSS) approach. A leader-follower cooperative learning strategy is put forward to mitigate the effects of noisy labels. To be specific, a leader network is first obtained based on transfer learning. Then, the leader network is jointly trained with two follower networks with the purpose of reducing the prediction divergence among the three networks. The small-loss criterion is employed to identify clean samples based on the joint loss function. A dynamic selection rate is introduced to automatically control the proportion of small-loss samples determined as clean during each epoch. The leader network, trained exclusively on the selected clean samples, is then utilized for extreme wind speed (EWS) prediction using real-world datasets. Furthermore, explainable artificial intelligence techniques are employed to improve the transparency and interpretability of the proposed TLACSS-based EWS prediction method.
How to cite: Liu, W., Wang, Z., Fang, J., Cao, Y., Liu, Y., Xue, Y., Salcedo-Sanz, S., and Liu, X.: Extreme Wind Speed Prediction Under Noisy Labels: A Transfer-Learning-Assisted Cooperative Sample Selection Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17880, https://doi.org/10.5194/egusphere-egu25-17880, 2025.
Flooding remains one of the most frequent and damaging natural disasters globally, exacerbated by climate changes and rapid urbanization. Understanding and mitigating urban flood risk requires near-real-time, fine-scale monitoring, including flood depth estimation. This study introduces a deep learning-based pipeline for estimating urban flood depth using device-independent street-level imagery, to complement existing remote, in-situ and hydrological approaches. By leveraging opportunistic sensing, this method exploits open-source tools to enhance spatial granularity and accessibility in flood monitoring.
The dataset, derived from a publicly available source, consisted of 3,367 annotated images of submerged vehicles, categorized into five flood levels based on water height relative to vehicle features (e.g., tires, chassis, windows). Cars were selected as reference objects due to their standardized dimensions and prevalence in urban environments, enabling consistent and reliable flood depth estimation.
The proposed pipeline processes images through an end-to-end workflow designed for real-time inference. It consists of four sequential stages: (1) vehicles are detected in street-level images using a pre-trained YOLO-World model; (2) detected vehicle regions are cropped and resized with a 20% bounding box enlargement to include flood visual indicators and additional context cues for the classification.; (3) images are super-resolved using pre-trained Enhanced Deep Super-Resolution (EDSR) networks to improve low-resolution imagery; (4) images are classified according the flood depth level using a ResNet50 model fine-tuned on the annotated dataset.
The classifier demonstrated robust performance across the five flood levels. The confusion matrix revealed minor misclassifications between adjacent classes, particularly Levels 0 and Level 1. One-vs-all area under the receiver operating characteristic curves (AUC) values exceeded 0.85 for all classes, with the highest performance observed for Level 4 (AUC = 0.98) and Level 0 (AUC = 0.94). Real-world validation using crowdsourced images from the 2021 flood in Central Europe confirmed the pipeline's reliability, delivering accurate and consistent flood level predictions in near-real-time.
This research advances urban flood monitoring by introducing a cost-effective and adaptable method for flood depth estimation that leverages existing devices without specialized hardware. The pipeline’s modular design ensures scalability and seamless integration into early warning systems and disaster response platforms. Future work will explore its application to aerial and drone imagery with oblique perspectives and develop cross-view geolocalization of flood depth measurements to improve spatial coverage and accuracy.
How to cite: Notarangelo, N., Wirion, C., and Van Winsen, F.: A Deep Learning Pipeline for Urban Flood Depth Estimation from Street-Level Imagery., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4320, https://doi.org/10.5194/egusphere-egu25-4320, 2025.
The Atlantic Meridional Overturning Circulation (AMOC) is a vital climate system component, transporting heat and influencing the stability of regional and global climate patterns. Recent research highlights its susceptibility to abrupt transitions driven by nonlinear feedback and external variability, underscoring the need for a probabilistic understanding of its dynamics.
The proposed framework incorporates stochastic forcing into a nonlinear deterministic box model to simulate climate noise, such as fluctuating freshwater fluxes and wind-driven variability (not necessarily with noise). This modification allows the model to capture a broader spectrum of AMOC behavior, including low-frequency oscillations, stochastic resonance, and regime shifts. The study will focus on the salinity advection feedback mechanism and its interaction with stochastic perturbations to determine probabilistic thresholds for AMOC stability under various climate scenarios.
We incorporate system identification techniques to further refine the stochastic box model used. Specifically, Langevin Regression is used to identify the stochastic nonlinear models that explain the observe hysteresis of the AMOC. Detailed probabilistic bifurcation diagrams that illustrate the AMOC’s sensitivity to stochastic forcing are obtained, thus facilitating the identification of critical parameters influencing regime transitions, and improving the understanding of the interplay between deterministic dynamics and external variability. The aim of these results are to refine predictive tools for assessing AMOC resilience to anthropogenic and natural climate forcings and provide insights into early warning signals for tipping points.
How to cite: Caamaño, A. J., del Arco-Fernández, E., Chidean, M. I., Salcedo-Sanz, S., and Casillas-Pérez, D.: Stochastic Box Modeling of AMOC: Variability, Thresholds, and Tipping Points, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20840, https://doi.org/10.5194/egusphere-egu25-20840, 2025.
Wildfires, intensified by shifting climate patterns, present a growing challenge globally. This contribution focuses on fire spread modeling as an approach to strengthen both prevention and response strategies. By combining physical modeling with ML optimized parameter optimization, fire spread simulations offer practical insights into fire behavior across diverse environmental scenarios. The capabilities are illustrated using three 3 example case studies across different regions and conditions: two in Athens, Greece, and one in the United States.
The Fire Propagation Simulation can be applied during ongoing events to anticipate the fire’s course and support timely interventions. It can also be used in hypothetical scenarios to assess the impact of prevention strategies and refine risk reduction plans.
The research addresses key challenges, including integrating firefighting tactics into simulations and overcoming uncertainties in environmental datasets. By incorporating multimodal datasets, this study aims to enhance our understanding of fire dynamics and offers actionable strategies for managing wildfire risks effectively.
How to cite: Wahbe, J., Gottfriedsen, J., Laux, D., Rovó, D., Ortman, E., Friend, J., Sarelli, A., and Liesenhoff, L.: Anticipating Wildfire Behavior: Fire Spread Modelling Case Studies in Greece and the U.S., EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17067, https://doi.org/10.5194/egusphere-egu25-17067, 2025.
It is widely recognised that the ability of measure Soil Water Content (SWC) is crucial to improve early warning systems for environmental hazards like floods, droughts, landslides, avalanches and wildfires. However, hydrological variables are notably more difficult to measure than meteorological variables. Common technologies to measure SWC are invasive point-scale probes, which are hardly representative of a wider area, unsuitable for coarse-textured soils and easy to be broken or lost. The main alternative is remote sensing, which suffers limits related to spatial resolution, measurement depth and continuity.
As an attempt to compensate for the lack of measurements of hydrological variables, computational models are widely used to derive them from meteorological ones. For example, the Canadian-developed Fire Weather Index (FWI) relies mainly on precipitations and temperature to evaluate the dryness of the soil. Indeed models still need to be validated and improved using measured data.
Proximal sensors based on the concept of Cosmic Rays Neutrons Sensing (CRNS) emerged as a reliable option for non-invasive measurement of SWC, within a large footprint (hectares), in depth (tens of cm) and with sub-daily resolution. CRNS is based on the detection of neutrons, which are generated in the atmosphere by the interaction of cosmic rays (high energy particles naturally flowing from space), then backscattered by the soil and effectively absorbed by water, due to their strong interaction with hydrogen. CRNS systems can easily be integrated in meteorological stations and operate autonomously also in remote areas, while transmitting the data for a real-time monitoring.
In the framework of the MOSAIC Project*, six CRNS systems manufactured by Finapp were installed in sites selected to span different altitudes vegetation types and exposures, integrating them into pre-existent meteorological stations. Computation of the FWI is also available for the same sites. We will compare the information provided by the CRNS with the output of the FWI and discuss how the model can be improved by integrating the SWC measurement.
*This work is part of the MOSAIC Project (Managing prOtective foreSt fAcIng clImate Change compound events), co-funded by the European Union through the Interreg Alpine Space programme (Project ID: ASP0100014), and it involves the use of data provided by courtesy of ARPAV (Dipartimento Regionale per la Sicurezza del Territorio).
How to cite: Del Savio, A., Gianessi, S., Zecchini, F., Rizzolo, R., Biasuzzi, B., Stevanato, L., Lunardon, M., and Gazzola, E.: Integrating the measurement of Soil Water Content by proximal Cosmic-Rays Neutron Sensors in the assessment of wildfire susceptibility, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5964, https://doi.org/10.5194/egusphere-egu25-5964, 2025.
Desert locusts are known as the world’s most destructive migratory pest. In the context of the European project EO4EU and European Space Agency (ESA) project IDEAS, a service has been developed, divided in two parts: a first part as early warning to monitor suitable ecosystem for locust to breed, a second part as impact assessment simulating the evolution of swarms.
The first part aims at predicting favorable breeding grounds for desert locusts seven days in advance by checking environmental conditions of the previous fifty days. Used environmental variables are Soil water content, Precipitation, and Temperature from ERA-5 land (Copernicus Climate). Additionally, NDVI (Normalized Difference Vegetation Index) from MODIS plays a role in the prediction. Locust information for model training was taken from a presence-only dataset provided by FAO’s Locust Watch. Actual most effective model is a customized version of Maxent. The latter is a statistical model widely used by researchers for species distribution modeling (SDM) as it is designed to work with presence-only datasets. Our model keeps Maxent's principles modifying the internal structure replacing the linear machine learning model with a Gated recurrent unit (GRU). This enables the model learning complex patterns and better understand the temporal evolution of features. Input data are time series where every time-step is a 5-day average of the above mentioned environmental variables, 50 days into 10 time steps. Data have been split into train and validation sets by using as training locust findings from 2000 to 2019 and as validation findings from 2020 to 2021. Since no locust absence information is provided, only two evaluation techniques are used: recall, which reaches 76%, and positively predicted area which is at ~17%.
The second part aims at evaluating the geographic footprint that adult locusts will have within a two-week time frame. The focus is on forecasting migration patterns, as locusts are able to travel long distances in short periods and explore new areas unpredictably. The strength of this model lies in its stochastic structure since it simulates an environmental-biased random walk on a 2D lattice, generating batches of diverse potential scenarios. This approach incorporates complex driving-factors for migrations and considers all various paths that swarms may take. Another strength is the ability to account for environmental conditions throughout the entire lifespan of desert locusts, enabling the prediction of future movements while also considering past ones. The model takes as input temperature and wind data while all the parameters and assumptions about the locust biology are taken from the FAO “Desert Locust Guidelines: Biology and Behavior”. Collecting environmental variables is essential, as they not only trigger migration events but also determine the direction and speed of swarm movement. Finally, the model produces output maps that estimate the probabilities of future appearance of swarms and their potential sizes.
Predicted results for both parts are showing promising correlation with FAO reports on desert locust activity, additionally ground verification are on-going in order to test the performance of the model.
How to cite: Houël, M., Grassi, A., Bellotto, K., and Azami, W.: Environmental plague monitoring : Desert Locust prediction with artificial intelligence and stochastic model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1720, https://doi.org/10.5194/egusphere-egu25-1720, 2025.
The economic impact of climate change on the humanitarian and disaster aid sector is escalating, with 2024 funding needs close to USD 50 billion and projections suggesting worsening conditions. The targets of this aid represent the most fragile countries in the world, but the number of people expected to be in need and the funding required to support them are unknown and difficult to assess – posing an information gap. This study will estimate the economic magnitude of climate impact for humanitarian assistance through 2080.
Leveraging machine learning and a modified damage function framework, we aim to model the relationship from top-down global temperature and precipitation variables, socioeconomic and vulnerability factors like GDP (gross domestic product) and HDI (Human Development Index) to bottom-up populations empirically exposed or affected by climate-related hazards, and finally to those requiring external aid to cope.
We apply Gaussian Process Regression (GPR), a learning method suitable for complex non-linear analysis, to explore this relationship for countries under various shared socioeconomic pathways (SSPs) and representative concentration pathways (RCPs). The unique data comprises the INFORM Risk and Climate Change indices as well as humanitarian datasets from the United Nations Office for the Coordination of Humanitarian Affairs (OCHA) and INFORM Severity. We hypothesise that the analysis will reveal increases in humanitarian needs driven by intensifying climate impacts and extreme events, with implications for resource allocation and policy priorities in the sector.
This novel solution addresses key gaps in the economic modelling of non-market climate risks and integrated assessments models (IAM), advancing the integration of people-based humanitarian data into climate impact assessments via machine learning. Concretely, it will quantify the human cost of a warming climate in the most vulnerable areas of the world and inform climate resilience financing on its priorities.
How to cite: Jäpölä, J.-P., Berlin, A., Fabri, C., Van Schoubroeck, S., Hrast Essenfelder, A., Marzi, S., Poljansek, K., Ronco, M., and Van Passel, S.: Future Costs of Climate Change for Humanitarian and Disaster Aid, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-10428, https://doi.org/10.5194/egusphere-egu25-10428, 2025.
Posters on site: Wed, 30 Apr, 16:15–18:00 | Hall X3
The simultaneous occurrence of high river discharges and storm surges represent a substantial hazard for many low-lying coastal areas.
Potential future changes in the frequency or intensity of such compound flood events is therefore of utmost importance.
To assess such changes large and consistent ensembles with storm surge and hydrological models are needed that are hardly available.
Often the occurrence of compound flood events is linked to the presence of certain atmospheric circulation types.
Future changes in the frequency of such patterns can be directly inferred from available climate simulations.
A frequently used classification of atmospheric circulation types are the so-called ‘Großwetterlagen’ by Hess and Brezowsky.
Here possible future changes in the occurrence of these ‘Großwetterlagen’ were analysed using data from 31 realisations of CMIP6 climate simulations for the emission scenarios SSP1-2.6, SSP3-7.0, and SSP5-8.5.
As the classification is subjective, a deep learning ensemble for the automatic classification was developed and applied.
In winter, a higher frequency of the atmospheric pattern Cyclonic Westerly towards 2100 could be inferred as a robust result among all models and scenarios.
As this circulation type is potentially associated with compound flooding in some parts of the European coasts, this points towards potentially increasing risks from compound flooding in the future.
How to cite: Heinrich, P., Hagemann, S., and Weisse, R.: Automated Classification of Atmospheric Circulation Types for Compound Flood Risk Assessment: CMIP6 Model Analysis Utilising a Deep Learning Ensemble, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-159, https://doi.org/10.5194/egusphere-egu25-159, 2025.
Urban flooding has occurred in many cities in recent years, causing economic and property losses in severe cases. Taiwan also faces the same risk of urban flooding. This study is divided into two stages. The first stage is to understand the possible causes of drainage bottlenecks in urban areas through urban hydrological and hydrological models, so as to understand the current urban drainage precipitation tolerance capacity to withstand rainwater. The second stage uses real-time and forecast rainfall to forecast water level are calculated by (1) automatic scheduling and calculation of the hydrological and hydraulic models and (2) artificial intelligence combined with big data to predict the water level as a disaster prevention warning and prediction tool. The forecast water levels are calculated for the bottleneck channel section, and then forecasts and other disaster prevention monitoring and early warning are combined with strategies such as early operation of the downstream outlet pumping station of the drainage system to improve the flooding problem in local low-lying areas. This study focuses on the problems faced by Taiwan in urban area drainage improvement projects, such as the complexity of traffic and underground pipelines for people's livelihood. Moreover, extreme rainfall caused by severe convective weather every summer has become one of the main causes of urban flooding. In addition, due to climate change, the hydrological characteristics of urban areas have also changed, such as two consecutive rounds of concentrated rainfall and multi-distributed rainfall. Concentrated rainfall in the region exceeded the drainage design protection standards and caused widespread flooding. In 2024, Taipei City, New Taipei City (located in northern Taiwan), and Kaohsiung City (located in southern Taiwan) all experienced two heavy rains, with rainfall exceeding 80-90 mm per hour, causing serious urban flooding disasters. This indicates that concentrated rainfall in the city makes it difficult for urban rainwater sewer drainage in local low-lying areas of the city to resist such floods. In addition to low-lying areas, it needs to further understand the hydrological disaster prevention in urban areas and which important urban roads are prone to urban flooding due to this type of rainfall, so as to provide disaster prevention strategies. Before the project is improved, the non-engineering rainwater system adaptation and disaster prevention strategies of urban drainage are carried out. According to the characteristics of drainage systems in different sections, suitable strategies are integrated to reduce the frequency of flooding in urban areas and improve the adaptation of urban precipitation tolerance capacity. Strategies such as precipitation tolerance capacity are formulated, and the case of New Taipei City is used to illustrate.
How to cite: Yang, S., Leu, W., Chen, M., Kuo, J., and Yeh, K.: Study on the precipitation tolerance capacity of urban drainage infrastructure and artificial intelligence disaster prevention and early warning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4918, https://doi.org/10.5194/egusphere-egu25-4918, 2025.
How to cite: Chidean, M. I., Casillas-Pérez, D., Caamaño, A. J., and Salcedo-Sanz, S.: Introducing L-Correlation for Climate Network Construction: Application to droughts analysis in the Iberian Peninsula, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19521, https://doi.org/10.5194/egusphere-egu25-19521, 2025.
Wind extremes, encompassing both high-intensity wind events and periods of diminished wind activity, pose multifaceted challenges across sectors such as renewable energy production, infrastructure resilience, and environmental risk management. These phenomena, driven by complex interactions within atmospheric systems, demand innovative analytical and predictive approaches. This study explores the application of artificial intelligence (AI) to address these challenges, focusing on its potential to enhance the identification of patterns, improve forecasting accuracy, and integrate diverse meteorological datasets. By leveraging machine learning models and exploring their adaptability to wind-related datasets, this work aims to outline a framework for robust analysis and prediction of wind extremes. The versatility of AI techniques in handling the complexities of wind extremes positions them as pivotal tools for improving preparedness and resilience in various sectors.
How to cite: Cornejo-Bueno, L., Peláez-Rodríguez, C., Guijo-Rubio, D., Marina, C., and Salcedo-Sanz, S.: Improving Resilience to Wind Extremes: An AI-Driven Approach, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3142, https://doi.org/10.5194/egusphere-egu25-3142, 2025.
Wind speed forecasting represents a significant challenge in the global transition to sustainable energy systems. Wind energy, characterised by zero greenhouse gas emissions and relatively low cost, is a renewable resource that depends heavily on meteorological conditions, which are inherently variable and unpredictable. This variability and intermittency present substantial obstacles to ensuring a consistent power supply, underscoring the importance of accurate wind speed prediction as a critical area of research. Among the various approaches explored to address this challenge, machine learning (ML) has emerged as a prominent solution. ML includes methodologies such as regression (predicting continuous values of wind speed) and nominal classification (predicting discrete categories of wind speed). In nominal classification, wind speeds are discretised into classes to provide essential information for wind farm operations. In this study, wind speeds are categorised into four classes: 1) very low speeds, 2) moderate speeds, 3) high speeds, and 4) extreme wind speeds. While both very low and extreme speeds result in no power generation, this work focuses on the extreme wind speed class, as these events often necessitate turbine shutdowns to prevent structural damage.
To address the challenges of wind speed forecasting with a focus on extreme wind events, we propose the use of ordinal classification, a ML paradigm specifically designed for tasks where output categories exhibit a natural order, as is the case in this work. This study evaluates hourly wind speed predictions for a wind farm in Spain, using data collected over more than 15 years. Additionally, input features include meteorological variables such as temperature, wind components (u and v), and sea level pressure, among others. Forecasts are generated for three time horizons (1h, 4h, and 8h) to provide sufficient lead time for mitigating risks associated with extreme wind conditions. Two ordinal classification models based on artificial neural networks (ANNs) are analysed: 1) an ANN coupled with the cumulative link model (CLM), and 2) an ANN using a soft labelling optimisation technique. Additionally, other competitive ordinal and nominal classification methods are included for comparative analysis.
The results demonstrate that the proposed models outperform a number of nominal and ordinal classification methods. The ANN coupled with CLM delivers superior overall performance across all four classes, while the ANN employing the soft labelling approach achieves higher accuracy in predicting extreme wind speed events. These findings underscore the potential of ordinal classification to enhance wind speed forecasting, contributing to more effective wind farm management and the broader integration of renewable energy sources.
How to cite: Guijo-Rubio, D., Gómez-Orellana, A. M., Vargas, V. M., Ayllón-Gavilán, R., Cornejo-Bueno, L., Moreno-Cano, F., Hervás-Martínez, C., Salcedo-Sanz, S., and Gutiérrez, P. A.: Wind speed prediction using ordinal classification: an analysis of extreme values, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15414, https://doi.org/10.5194/egusphere-egu25-15414, 2025.
This study introduces an innovative preprocessing technique utilizing an Autoencoder (AE) as an alternative to the traditional multivariate Analogue Method (AM). The newly proposed method, MvAE-AM, is employed to reconstruct historical heat wave events: France in 2003, the Balkans in 2007, Russia in 2010, and Spain in 1995. The AE effectively extracts critical information from variables such as soil moisture (SM), potential evaporation (PEva), mean sea level pressure (MSL), and geopotential height at 500 hPa (Z500) into a more compact univariate latent space. Subsequently, the conventional univariate AM is utilized to identify analogous past situations within this latent space, focusing on minimizing the distance to the analyzed heat wave. This analysis is extended to comparing factual and contrafactual scenarios, where the attribution of the anthropogenic impact can be studied. Our evaluation of the proposed MvAE-AM method against the standard multivariate AM (MvAM) reveals that it not only simplifies the complexity of the problem but also enhances accuracy. Furthermore, a significant advantage of the AE-based approach over classical statistical methods is its capacity for detailed explainability analysis, facilitated by explainable artificial intelligence (XAI) techniques such as SHAP. This analysis elucidates the temporal, spatial, and variable-specific factors that most significantly influence heat wave occurrences, with notable patterns of Ridges and Blocking observed across several heat wave events.
How to cite: Marina, C. M., Lorente-Ramos, E., Cornejo-Bueno, L., Barriopedro, D., García-Herrera, R., Giuliani, M., Scoccimarro, E., Zorita, E., Castelletti, A., and Salcedo-Sanz, S.: Attribution with Multivariate Analogues: a heat waves scenario, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3203, https://doi.org/10.5194/egusphere-egu25-3203, 2025.
Extreme climate events like storms, heatwaves, wildfires, floods, and droughts pose serious threats to human society and ecosystems. Measuring their impacts remains a crucial challenge scientifically. Although data linking climate hazards to socio-economic effects are crucial, their public availability is still relatively sparse. Existing open databases such as the Emergency Events Database (EM-DAT) and DesInventar Sendai offer some impact data on climate extremes, but impact data on climate extremes also appear in newspapers, reports, and online sources like Wikipedia.
We introduce Wikimpacts 1.0, a comprehensive global database on climate impacts developed using natural language processing techniques. This database utilizes the GPT4o large language model for extracting information, following document selection, post-processing, and data consolidation. In this release, we have processed 3,368 Wikipedia articles. Impact data for each event is recorded at three levels: event, national, and sub-national. Categories include the number of deaths, injuries, homelessness, displacements, affected individuals, damaged buildings, and insured or total economic damages. This dataset encompasses 2,928 events from 1034 to 2024, featuring 20,186 national and 36,394 sub-national data entries. Comparison with manually annotated data from 156 events shows that the Wikimpacts database is highly accurate in the event level for time, location, deaths, and economic damage, though details on injuries, affected individuals, homelessness, displacements, and building damage are slightly less precise. An analysis from 1900 to 2024 demonstrates that sub-national data provides more comprehensive coverage of tropical and extratropical storms, and wildfires than EM-DAT, with enhanced data on events in countries like the United States, Mexico, Canada, and Australia. Our study emphasizes the potential of natural language processing in creating open databases with reliable information on climate event impacts.
How to cite: Li, N., Thiery, W., Zahra, S., Madruga de Brito, M., Worou, K., Kurfali, M., Lampe, S., Munoz, P., Flynn, C., Trigoso, C., Nivre, J., Zscheischler, J., and Messori, G.: Wikimpacts 1.0: A new global climate impact database based on automated information extraction from Wikipedia, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8954, https://doi.org/10.5194/egusphere-egu25-8954, 2025.
Climate change is driving an alarming rise in extreme weather events, including heatwaves, droughts, and floods. Among these, heatwaves stand out as the deadliest, with profound and widespread impacts across multiple sectors. Europe is emerging as a global heatwave hotspot, with heatwave frequency increasing almost four times faster than other northern midlatitudes. Agriculture is the most vulnerable sector to temperature extremes, making adaptation measures essential to support food security worldwide.
In this work, we investigate the impacts of temperature extremes on agricultural productivity in the Adda River basin in northern Italy, in order to inform the design of adaptation strategies and to respond to projected mid-to-long-term climate change. We first simulate historical crop yields using a detailed, process-based model of the agricultural districts. Then, we use correlation analysis and the Patient Rule Induction Method to identify key drivers of crop failure. These drivers include various indices that quantify the occurrence and intensity of heatwave and drought events. Numerical results suggest that the two most important indices are the number of days above the climatology 90th percentile (NDQ90) in June, calculated with daily maximum temperature, and the nighttime Heat Wave Magnitude Index (HWMI), calculated with daily minimum temperature.
Lastly, we evaluate the projected evolution of these two indices using six CMIP6 climate models across four climate change scenarios. To integrate climate information independently of specific scenarios, models, or periods, we analyze the ensemble of future projections by focusing on two Global Warming Levels (GWLs) calculated with respect to each model’s pre-industrial global temperature, 1.5 °C and 4.0 °C. These are compared to a baseline at a GWL of 0.69 °C corresponding to the warming level of the climatology used to compute the indices from reanalysis data. Our results show that, while both indices are projected to grow considerably relative to the reference period, HWMI displays the greatest increments, with an ensemble average that increases 21-fold when moving from GWL 0.69°C to 4.0 °C. For NDQ90 in June this variation is from 3.40 to 17.09, indicating that, on average, more than half the days in June will experience extreme maximum temperatures at GWL 4.0 °C. These trends suggest the opportunity to replace some of the crops currently cultivated in the area, primarily maize, with more heat-tolerant varieties, such as soy or cereals, in order to ensure a more reliable production in the coming years and decades.
How to cite: Calisi, C., Giuliani, M., McAdam, R., Squintu, A., Scoccimarro, E., and Castelletti, A.: Assessing the impacts of temperature extremes on crop production in the Adda River basin , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11027, https://doi.org/10.5194/egusphere-egu25-11027, 2025.
Posters virtual: Wed, 30 Apr, 14:00–15:45 | vPoster spot 3
EGU25-14209 | Posters virtual | VPS13
A digital twin for management of landslides and slope incidents on strategic road infrastructureWed, 30 Apr, 14:00–15:45 (CEST) | vP3.17
In 2023, Otis strengthened from a slight tropical storm into a major hurricane (Category 5) within only about 12 hours before it made landfall. The storm slammed into Mexico's coast with maximum sustained winds of over 165 mph and hurricane-force winds extending up to 30 miles from its center. The SICT (Secretariat of Infrastructure, Communications and Transportation) warned of a total closure of the Mexico-Acapulco highway in the Chilpancingo-Acapulco section. Faced with reports of hundreds of landslides through the lines, the SICT deployed more than 1000 workers, 100 vehicles and 300 pieces of heavy machinery in the hope of “restoring traffic as soon as possible and providing safety to users.” Unfortunately, predictions could not anticipate close enough the Otis destructive force.
Ensuring the proper functioning of road infrastructure is a fundamental aspect in risk management. Landslides have the potential to impair critical transportation infrastructure, particularly road networks in the hilly regions in Mexico. Recognizing the extremely changing climate conditions in the Mexican Pacific coasts are becoming increasingly difficult to predict, in this research advanced technologies are integrated into an intelligent digital scenario to simulate and control this linear infrastructure before, during and after extreme rainfalls occur.
The strategic roads digital twin comprises i. dynamic susceptibility maps, ii. satellite radar information of control points (the landslides pathologies are easily detected through them), iii. an artificial intelligence slope stability calculator (in near real-time) for pointing incipient instability, and iv. a semi-immersive scenario for analyzing future states based on the information of pluvial stations and control points, once this information is analyzed with the intelligent calculator. For communicate the input conditions, the aggravating factors and the future responses, a digital twin of potentially affected road sections (detected on the dynamic maps) is developed. Simulate scenarios before rainfall increases, help to make informed maintenance and risk prevention decisions in road infrastructure in areas with high geotechnical complexity and strong seasonal rainfall patterns. Exploiting precalculated extremely dangerous conditions, this digital twin can serve as an early warning system because it is programmed for immediate communication of graduated alarms that announce the proximity to dangerous states.
How to cite: García, S., Trejo, P., and Ángeles, B.: A digital twin for management of landslides and slope incidents on strategic road infrastructure, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14209, https://doi.org/10.5194/egusphere-egu25-14209, 2025.
