AI for Extreme Event Modeling and Understanding: Methodologies and Challenges

Abstract

In recent years, artificial intelligence (AI) has deeply impacted various fields, including Earth system sciences. Here, AI improved weather forecasting, model emulation, parameter estimation, and the prediction of extreme events. However, the latter comes with specific challenges, such as develo** accurate predictors from noisy, heterogeneous and limited annotated data. This paper reviews how AI is being used to analyze extreme events (like floods, droughts, wildfires and heatwaves), highlighting the importance of creating accurate, transparent, and reliable AI models. We discuss the hurdles of dealing with limited data, integrating information in real-time, deploying models, and making them understandable, all crucial for gaining the trust of stakeholders and meeting regulatory needs. We provide an overview of how AI can help identify and explain extreme events more effectively, improving disaster response and communication. We emphasize the need for collaboration across different fields to create AI solutions that are practical, understandable, and trustworthy for analyzing and predicting extreme events. Such collaborative efforts aim to enhance disaster readiness and disaster risk reduction.

1 Introduction

The frequency, intensity, and duration of weather and climate extremes have increased in recent years, posing unprecedented challenges to societal stability, economic security, biodiversity loss, and ecological integrity seneviratne2021weather . These events –ranging from severe storms and floods to droughts and heatwaves– exert profound impacts on human livelihoods and the natural environment, often with long-lasting and sometimes irreversible consequences. Modeling, characterizing, and understanding extreme events is key to advancing mitigation and adaptation strategies.

In this context, Artificial Intelligence (AI) has emerged as a transformative tool for the detection aidetoolbox , forecasting Lam2023 ; ferchichi2022 , analysis of extreme events and generation of worst-case events ragone2021rare , and promises advances in attribution studies, explanation, and communication of risk Yokota2004 . The capabilities of machine learning (ML), and deep learning (DL) in particular, in combination with computer vision techniques, are advancing the detection and localization of events by exploiting climate data, such as reanalysis and observations salcedosanz2024analysis . Modern techniques for quantifying uncertainty gawlikowski2023survey are a necessary step to making progress in climate change risk assessment harrington2021quantifying . Using ensembles in combination with AI models has progressed the field of attribution of extremes stott2016attribution , identification of patterns, trends madakumbura2021anthropogenic and climate analogs chattopadhyay_analog_2020 . Yet, AI does not only excel at prediction but can also explain processes (e.g., via explainable AI FindingtherightXAIMethodBommer2024 and causal inference Hannart2016CausalCT ), which is essential for decision-making and effective mitigation strategies. Recent advances in large language models (LLMs) that retrieve information from heterogeneous text sources allow for effective integration of methods for communication tasks and human-machine interaction in extreme event and situation analysis vaghefi2023chatclimate .

Figure 1: A general AI-driven extreme event analysis pipeline. Different components in modeling and understanding extreme events using AI methodologies are interconnected, highlighting the flow from data collection and analysis to actionable insights/outputs and the challenges encountered. Note the feedback loops where AI does not only produce some relevant outputs and products from data (predictions, patterns and trends, climate attributions, and causal relations) but also may help suggest areas for improvement and adaptation in methodologies to overcome identified challenges, redefine the hypotheses and challenges themselves, as well as inform data collection and preprocessing.

Here, w review the role of AI for extreme event analysis, its challenges and opportunities for extreme event analysis. The general pipeline of AI-driven extreme event analysis (cf. Fig. 1) encapsulates the entire workflow from data collection and preprocessing to the generation of outputs such as predictions, patterns, trends, climate attributions, and causal relations. The framework also highlights the iterative nature of AI processes, where outputs serve their direct purpose and also serve to inform and improve data collection, preprocessing, and hypothesis formulation. Nevertheless, develo** AI models that are accurate, operational, explainable, and trustworthy still remains a significant challenge. Issues such as data scarcity, the need for high-dimensional data processing, and the interpretation and explanation of AI models are critical challenges that must be addressed. Moreover, communicating risk effectively and ethically, and integrating AI tools with existing systems to ensure operational viability and scientific robustness are nontrivial tasks requiring ongoing research and development.

We provide a comprehensive overview of the AI-driven pipeline for analyzing extreme events, including modeling, detection, forecasting, and communication (Section 2), and discuss key challenges (Section 3). We present case studies on AI applications for droughts, heatwaves, wildfires, and floods (Section 4), and conclude by identifying major challenges and future research opportunities in AI for extreme events (Section 5).

2 Review of AI methods

This section reviews the main methods in all aspects of extreme event analysis (cf. Table 1): data, modeling, understanding and trustworthiness, and the last mile on communication and deployment.

Table 1: A comprehensive categorization of extreme events analysis in four main blocks (data, modeling, understanding and trustworthiness, and the last mile on communication and deployment) regarding their task, standard and current AI methods.

2.1 Extreme Event Modeling

AI methodologies for extreme event modeling can be categorized into detection ruff2021unifying , prediction han2019review and impact assessment zennaro2021exploring tasks. Thanks to the emergence and success of DL, all of these tasks can be tackled by designing data-driven models that exploit spatio-temporal and multisource Earth data characteristics, from climate variables to in situ measurements and satellite remote sensing images (see Figure 2 [top row]).

2.1.1 Detection

Detecting and identifying extreme events geographically over time to distinguish them from normal conditions is fundamental for discovering underlying patterns and correlations between climate variables. This allows for a better understanding of their generating processes and mechanisms and facilitates their anticipation.

Classical statistical methods, such as threshold or percentile-based indices, have been widely applied to detect extreme events (see Table 1). However, these often fail to identify the same events as expert or impact-based detection methods Jones_2023 . AI methodologies can help reconcile expert knowledge with data-driven approaches as they capture regularities and nuanced relations in large volumes of observational data. Canonical ML treats the detection problem as a one-class problem or an outlier detection problem. Many methods are thus applied flach2017multivariate , and available in software packages aidetoolbox . Recent advances include deep learning for segmentation and detection of tropical cyclones (TCs) and atmospheric rivers (ARs) in high-resolution climate model output karaismailoglu2021climatenet and semi-supervised localization of extremes racah2017extremeweather . Alternatively, reconstruction-based models (e.g., with autoencoders) are optimized to accurately reconstruct normal data instances, and thus extremes are associated with large reconstruction errors GuancheGarca2018 . Finally, probabilistic approaches try to identify extremes by estimating the data probability density function (PDF) or specific quantiles thereof. Standard Extreme value theory (EVT) approaches, however, struggle with short time-series data, nonlinearities, and nonstationary processes. Alternative probabilistic (ML) models have relied on surrogate meteorological and hydrological seasonal re-forecasts Klehmet2024 , Gaussian processes and nonlinear dependence measures Johnson20sakame and multivariate Gaussianization Johnson20rbig4eo .

2.1.2 Prediction

Designing predictive systems that accurately model extreme events is essential to anticipating the effects of future extreme events and providing critical information for decision-makers to prevent damages. Spatial and temporal predictions aim to provide a quantitative estimate of the future value of the Earth’s state (Table 1).

Many ML algorithms have been proposed for deterministic extreme event prediction, but most are only applied to small regions and specific use cases, though. Prediction can be performed using climate variables alone vijverberg2022role or in combination with satellite imagery kladny2024enhanced . Here, for example, the vegetation response under extreme drought conditions is modeled as a video prediction task of the satellite imagery while conditioning on the climate variables. Another common approach is to directly estimate an indicator that defines the extreme event, e.g. flood risk maps bentivoglio2022deep or drought indices belayneh2014long . The estimated variables can cover different lead times depending on the characteristics of the extreme, spanning from short-term prediction to seasonal prediction kondylatos_wildfire_2022 . Recently, DL-based prediction techniques have gained popularity for their ability to process large data volumes, capture complex nonlinear relationships, and reduce manual feature engineering. These benefits have led to the creation of global models that generalize across various locations, as seen in flood nearing2024global and wildfire zhang2021deep predictions.

Probabilistic models, unlike deterministic ones, focus on predicting the probability distribution of future variable states. The importance of probabilistic forecasts for extreme heatwaves has been emphasized miloshevich2023probabilistic . AI-based techniques, integrated within climate models, can enhance predictions, such as in drought prediction vo2023lstm and extreme convective precipitations downscaling shi2020enabling .

2.1.3 Impact Assessment

Estimating the effects of extreme events on society, the economy, and the environment is crucial for conveying potential future consequences to the public, policymakers, and across disciplines callaghan2021machine ; zennaro2021exploring . Impact assessment involves understanding how a system reacts to extreme event forcings. Unlike extreme event detection and prediction, the focus here is on impact-related outcomes, such as the number of injuries, households affected, or crop losses.

In recent years, there has been an increasing interest in predicting vegetation state using ML ferchichi2022 , as a way to estimate the impact of climatic extremes on the evolution of the vegetation state variable sutanto2019moving ; salakpi2022forecasting . Recent advances methods have used echo state networks martinuzzi2023learning , ConvLSTM-based ahmad2023machine ; kladny2024enhanced , and transformers benson2023forecasting using high-resolution remote sensing and climate data.

Another way to address impact assessment with ML is to analyze changes in the probability density function over time. This approach allows a quantification of the impact of different events, which can be used to improve our understanding of the drivers of vulnerability, zhang2021impact to such as population displacement ronco2023exploring . Alternatively, the impact of extreme events can be detected by analyzing the news coverage based on natural language processing (NLP) and, more recently, on large language models (LLMs) Sodoge2023 .

Figure 2: Components in an AI pipeline for extreme events. AI mainly exploits spatio-temporal Earth observation, reanalysis, and climate data to answer “what’-questions (top row): detection of events, prediction, and impact assessment. AI can also be used for understanding events and thus answer “what if,” “why,” and “how sure” questions (middle row) and makes use of explainable AI (XAI) to identify relevant drivers of events, causality to understanding the system, estimate causal effects and impacts, and imagine counterfactual scenarios for attribution and uncertainty estimation to quantify trust and robustness for decision-making. Communicating extreme events and their impacts can benefit from statistical/machine learning (bottom row) by improving operationalization, ensuring fair and equitable narratives, and integrating language models in situation rooms for enhanced decision-making.

2.2 Extreme events understanding and trustworthiness

Yet, all previous approaches focus on the ‘what,’ ‘when,’ and ‘where’ questions, not the ‘why,’ ‘what if,’ and ‘how certain’ ones. In environmental sciences, there’s an increasing emphasis on the need for trustworthy machine learning bostromTrustTrustworthy . This is especially important in extreme events, where critical high-stakes decisions need to be made impacting public safety, health, infrastructure, and resource allocation. Disciplines like explainable AI (xAI) and uncertainty quantification (UQ) offer methods to make AI more reliable and trustworthy (see Figure 2 [middle row]). These approaches not only help us interpret AI model predictions but also enhance our understanding of extreme events themselves. Techniques such as causal inference and extreme event attribution further enhance xAI and UQ by understanding the mechanisms behind these events, crucial for improving AI models and gaining trust in the decision-making process.

2.2.1 Extreme Event Attribution

Extreme event attribution (EEA) quantifies the influence of anthropogenic forcings (such as greenhouse gas emissions) on the likelihood of extreme climate events by using numerical simulations with General Circulation Models (GCM) to compare their probabilities under observed conditions (the factual world) and a hypothetical scenario without human emissions (the counterfactual world) naveau2020statistical (cf. Table 1). Two main viewpoints exist: probabilistic EEA employs quantitative statistical methods to estimate this likelihood (e.g., the EE is 600 times more likely due to human emissions), while storyline approaches simulate the evolution of the EE under different forcings to gather a process-based attribution statement (e.g., 50% of the magnitude of an EE is explained by natural variability) otto2023attribution . Ideally, both methods should be combined to provide a complete understanding.

Ensembles of neural networks have been developed to emulate GCMs, creating surrogate models that operate significantly faster pasini2017attribution . These models successfully predict the year based on the global annual temperature field under present conditions but struggle to perform accurately in hypothetical scenarios resembling pre-industrial conditions barnes2019viewing . Despite the scarce literature on this topic, the impressive recent advances in weather and climate emulation with AI Lam2023 suggest that climate emulators will play a significant role shortly, allowing almost real-time EEA.

2.2.2 Explainable AI

Many models are transparent and interpretable-by-design, like linear models or decision trees, but they may not perform well on complex problems and are rarely employed in extreme events (cf. Table 1). Explainable AI (xAI) aims to unveil the decision-making process of AI models. xAI also facilitates debugging, improving models, and gathering scientific insight by revealing the model functioning, learned relationships, and biases. The most commonly used xAI approaches rely on model-agnostic distillation or feature attribution methods ghaffarianExplainableArtificial2023 ; Tuia24perspective , cf. Fig. 2. Distillation methods, such as SHapley Additive exPlanations (SHAP) and Local interpretable model-agnostic explanations (LIME), create surrogate models and have been widely used in geosciences and climate sciences schlund2020constraining ; ronco2023exploring . Feature attribution methods, like Partial Dependence Plot (PDP) or Gradient-based Class Activation Map (Grad-CAM), highlight important features by perturbing the inputs or using backpropagation srinivasan2020machine . Recent approaches explain DL models using attention in drought prediction dikshitSolvingTransparency2022a and prototypes for explaining event localization barnesThisLooks2022 . xAI have been also used to evaluate climate predictions FindingtherightXAIMethodBommer2024 ; mamalakisInvestigatingFidelity2022 , and thus offering an agnostic, data-driven approach for model-data intercomparison.

2.2.3 Causal inference

Causal graphs encode intuitive information about robust system properties. This information is further required for deciding the identifiability of queries about interventional or counterfactual properties (pearl2017causality, ; Peters2017elements, ). For extreme events, most widespread causal inference methods Runge2023 ; CampsValls23physcausaldiscovery cannot be applied directly, as traditional assumptions of normality and mild outliers do not hold, and dependence may manifest only in the extremes (Table 1). This has motivated recent works that deal with causal discovery for extreme event analysis, providing different frameworks for understanding the dependencies and causal structures in extreme values Gnecco2019CausalDI , with applications on, e.g. rivers discharge Pasche2021CausalMO . Another relevant line of work focuses on answering counterfactual questions for climate extreme events, which links to EEA (Hannart2016CausalCT, ; Kiriliouk2019ClimateEE, ).

2.2.4 Uncertainty quantification

Even equipped with explanatory and causal methods, it is still crucial to assess how confident are AI model decisions, as inaccurate warnings or decisions could impact safety and resources (gawlikowski2023survey, ; ghanem2017, ). Understanding the sources of uncertainty is important to inform and disentangle the inherent (aleatoric) uncertainty of the weather phenomenon from the lack of knowledge in the model (epistemic) uncertainty. One can reduce the latter with more data and additional assumptions, but not the former (cf. Table 1). DL approaches with UQ have shown promise for extreme events recently miloshevich2023probabilistic ; xu2022quantifying .

2.3 The Last Mile: operationalization, communication, ethics and decision-making

The previous components (methods and techniques) in the AI pipeline need to be operationalized, be robust, accountable and fair, so they ultimately serve the purpose of evidence-based policy making (see Figure 2 [bottom row]).

2.3.1 Operationalization

Operationalization strictly needs the previous layer in place (xAI, causality, UQ) for improved and accountable explanations of the predictions in early warning systems (EWSs), and for enhanced disaster risk management. Calibration of ML models bella2010calibration , especially when coupled with uncertainty estimation, aligns predicted probabilities to reflect the real-world likelihood of extreme outcomes, enhancing reliability and interpretability essential for informed decision-making. Poor calibration can lead to misguided decisions, such as overestimating or underestimating the likelihood of critical events.

2.3.2 Communication of risk and ethical aspects

History shows that even actionable forecasts can fail if not communicated properly. For instance, despite predictions of the Mediterranean storm Daniel’s landfall four days in advance, the lack of effective communication contributed to the tragic outcome in Libya, with severe casualties and displacement corps2023libya . Even in the more developed country of Germany, the devastating effects of floods were related to ineffective warnings dwd2021weather . This highlights the critical need for robust EWSs that predict events and effectively communicate risks to ensure community preparedness and response (Table 1).

The challenge of false alarms in the context of EWS is a significant challenge, as they can lead to “warning fatigue”, where the public becomes desensitized to alerts and may ignore crucial warnings during actual emergencies tojcic2021performance ; yore2020early ; reichstein2023ews . Addressing this issue requires improving the accuracy and trust in predictive models, refining communication strategies, and engaging the community tamamadin2020automation .

2.3.3 Ethics in AI models and data

The governance of AI ethics calls for systems to be respectful of human dignity, ensure security, and support democratic values RNY004 . The deployment of AI to help manage extremes involves several fundamental principles: ensuring fairness, maintaining privacy, and achieving transparency RNY003 ; kochupillai2022earth . In this context, the rise of Large Language Models (LLMs) heightens ethical risks like outdated or inaccurate data, bias, errors, and misinformation, often obscured by their vast training on web data. Additionally, current AI models (like generative AI and LLMs) often draw on large datasets, so there is a risk of strong biases if these datasets are unrepresentative Tuia24perspective . Spatial sampling and analysis are vital to collecting geographically and environmentally representative, fair, and unbiased data. Global initiatives like the Common Alerting Protocol have standardized warning data, but the success of these systems depends on their inclusivity and ability to adapt to the diverse needs of affected communities, who should participate in develo** EWSs macherera2016review ; reichstein2023ews . AI enhances these systems by enabling the rapid dissemination of personalized alerts tailored to specific locations and individual risk factors like proximity to flood plains or wildfire zones, ensuring warnings are understandable and relevant to all.

Achieving the inclusion of affected communities, especially in the Global South, still remains a significant challenge, as centralized, ’one-size-fits-all’ models are often simpler and cheaper to implement but less effective at addressing local nuances. Designing AI systems, building on large-language models fine-tuned to local communities with users in the loop, offers excellent opportunities to overcome “one-size-fits-all” while being efficient.

2.3.4 Policy and decision-making

Even if AI-assisted, human operators are in charge of implementing the final decision. This implies supporting end-users in making the most out of the generated information Yokota2004 . This operational value is often problem-specific and depends upon the system’s dominant dynamics and the socio-economic sectors considered. Quantifying this value, however, might not be straightforward as a more accurate prediction does not necessarily imply a better decision by the end users Ramos2013 . When multiple forecasts from different systems are available, users should address a number of problems where AI can be helpful, including the selection of the forecast product, the lead time, the variable aggregation, the bias correction, and how to cope with the forecast uncertainty.

Integrating the pipeline in Fig. 1 with impact models that leverage Reinforcement Learning algorithms to simulate optimal decision-making can help quantify how AI-enhanced information about extreme events translates into better decisions Giuliani2015ISA . The co-design of impact models via participatory processes, including end users in the loop, further strengthens the overall model-based investigation by better capturing end users’ requirements, expectations and concerns. End users should understand both the benefits of AI-enhanced information for improved decision-making and the risks of misinformation leading to poor choices.

3 Data, model and integration challenges

Extreme event analysis faces many important challenges related to the data and model characteristics, but also to the integration of AI in decision pipelines (see Table 2.

Table 2: Challenges and risks of using AI for extreme event analysis, and needed advances in AI to tackle them. The same blocks before related to data, model and integration challenges are studied.

3.1 Data challenges

A major challenge is the lack of sufficient data with expert annotations, which are essential for training and evaluating AI models. Given their rarity, extreme events can be overlooked during the data preprocessing steps to eliminate noise, gaps, biases, and inconsistencies camps2021deep ; Reichstein19nat . Additionally, AI struggles with integrating and extracting relevant information across various data sources and scales, which can complicate feature extraction and selection Tuia24perspective . Future AI development needs to focus on deriving robust features (or representations) that effectively capture the distinct characteristics of extreme events (cf. Table 2).

AI models are increasingly used to enhance parameterizations of subgrid-scale processes in Earth system models, addressing gaps in traditional methods. However, a key challenge is their numerical instability over extended time horizons, which may produce unrealistic scenarios when simulating extremes due to insufficient training data iles2020benefits . The quality of observational data also challenges AI methods used for data assimilation. Recently, hybrid ML models that combine domain-driven with data-driven models approaches promise more robust and trustworthy AI models cheng2023machine ; Reichstein19nat . Moreover, AI-based methods have improved error characterization, enhancing uncertainty quantification Haynes2023 . Generative models are also used to sample ensemble members more effectively, offering better representations of the system’s state Li2022 (cf. Table 2).

3.2 Model challenges

The lack of a clear statistical definition of extreme events and the mechanisms responsible for their occurrence hamper model development and adoption. For detection, extremes usually constitute not pointwise but complex contextual, group, or conditional anomalies, whose source (changing process, parents, distribution) is often unknown Tuia24perspective . This results in challenges, such as capturing subtle (new) patterns, setting adaptive thresholds, or integrating data across distant points in space and time (cf. Table 2). For prediction and impact assessment, AI models are sensitive to initial conditions and may not capture long-term dependencies camps2021deep . Moreover, data might not reveal the dynamics of extremes, there might be changes to unseen dynamics flach2017multivariate , and stationarity may not hold in general. Using hybrid models that combine data, domain knowledge and models could allow insights into the mechanisms that trigger extreme events Reichstein19nat ; farazmand2019extreme ; cheng2023machine .

The complexity of extremes also makes attribution, causal discovery, and explainability particularly challenging. xAI can only reveal correlations the model learned and has no information about the causal structure. This could lead xAI to exacerbate model biases or spurious correlations. Indeed, different xAI methods can produce very different explanations, whose suitability for different models should be quantitatively evaluated FindingtherightXAIMethodBommer2024 . Causality is not error-free, as a wrong assumption about where the extreme comes from could lead to wrong causal graphs, conclusions and decisions pearl2017causality . Finally, challenges in UQ include differentiating and quantifying aleatoric and epistemic uncertainties, which is complicated by the models’ overparameterization and their lack of robust probabilistic foundations (cf. Table 2).

3.3 Integration challenges

ML models are typically trained on high-quality, well-curated datasets, such as Copernicus ERA 5 reanalysis or cloud-free satellite imagery, which often do not reflect the error-prone meteorological forecasts and cloudy conditions encountered in real-world situations (Table 2). Applying domain adaptation strategies or leveraging invariant features could align model performance from training phases to operational conditions. Besides, leveraging proprietary and trusted geospatial data from operational stakeholders, such as detailed forest fuel maps and elevation models, allows for fine-tuning these models to enhance detection and forecasting accuracy and enable finer spatial and temporal resolution in output products.

4 Case Studies

We showcase four case studies on drought, heatwaves, wildfires, and floods, each of them covering the different aspects of detection, forecasting, impact, explainability, attribution, and communication of risk (Fig. 3).

Figure 3: Summary of case studies using AI to manage extreme events. Four use cases (drought, heatwaves, wildfires, and floods) are showcased where AI enables detection, forecasting, impact assessment, explanation, understanding, and communication of risk, providing a comprehensive solution for disaster management. (a) Droughts. Top: AI leverages multimodal data to predict Earth’s surface dynamics, enhancing forecasts for crop yields, forest health, and drought impacts. Bottom: XAI techniques, like “neuron integrated gradients,” elucidate the key factors driving severe drought conditions, highlighting variable interactions over time. (b) Heatwaves. Top: Variables of interest are extracted from heterogeneous data sources (images, time series, text) and potentially aggregated over space and/or time. Bottom Left: Relevant features can be extracted from the data using e.g., clustering techniques. Right: Prediction of heatwaves can be done in combination with dimensionality reduction tools or directly from the selected features. (c) Wildfires. Top: AI enhances understanding and prediction of wildfire dynamics, particularly for mega-fires intensified by global warming, by analyzing extensive datasets and differentiating fire types with XAI. Bottom: AI combined with causal inference aims to better detect and understand pyrocumulonimbus clouds, intense storm systems generated by large wildfires that complicate fire behavior prediction. (d) Floods. AI transforms flood risk communication by using realistic 3D visualizations and animations to depict rising water levels’ impact on communities and infrastructure, making the information more relatable (thisclimatedoesnotexist.com). AI-driven platforms analyze vast amounts of data from weather forecasts, river levels, and historical flood patterns to predict future events accurately, integrating this information with digital maps and urban models to identify high-risk areas (climate-viz.github.io.com). This approach enhances flood risk management by allowing for targeted, personalized communication, enabling residents to receive specific alerts and visualize potential impacts on their homes. AI also supports the generation of detailed flood reports from various sources, enhancing preparedness and mitigation efforts (floodbrain.com).

4.1 Droughts: detection, forecasting and explainability

Droughts are among the costliest natural hazards, having destructive effects on the ecological environment, agricultural production, and socioeconomic conditions. They have been traditionally approached with heuristic thresholds and simple parametric models on related variables (e.g., soil moisture, vegetation indices, precipitation, or temperature, etc.). Being drought monitoring historically purely based on station-based measurements, Earth observation technology (e.g., Sentinel and Landsat series) and remote sensing allowed for estimating drought-related variables over larger spatial and temporal scales west2019remote . However, drought definitions presented in the literature (meteorological, hydrological, agricultural, operational, or socioeconomic) are usually subjective and limited zargar2011review . These types are not independent but refer to complementary physical, chemical, and biological processes of the Earth system droughts involve. Although empirical multivariate drought indicators have been proposed to account for these dependencies, the complexity of the combination of drought processes makes their detection, prediction, and characterization (severity, duration, start-end, etc.) a burden to researchers and policymakers, which can benefit from data-driven AI approaches.

Traditional ML approaches have been successfully applied to drought prediction, from support vector machines, decision trees, and random forests to more advanced multivariate density estimation approaches jehanzaib2021comprehensive ; 2024mengxue ; Johnson20rbig4eo . However, DL algorithms are required for efficiently modeling spatial and temporal correlations in data. To this end, neural networks have been used for drought monitoring in a supervised way, mainly using multilayer, convolutional, and/or recurrent neural networks aidetoolbox . Drought detection is an unsupervised problem due to its spatiotemporal variability and the shifting definition of drought under climate change. Thus, unsupervised and semi-supervised methods are recommended for practical applications ruff2021unifying .

To overcome some of the challenges before, recent works have applied AI methods to process multimodal data (satellite imagery at different resolutions, mesoscale climate variables, and static features). Drought detection has been addressed at the European level employing domain knowledge-driven variational autoencoders (VAEs) 2024mengxue , which combine traditional drought indicators with climate data. Forecasting the dynamics of the Earth’s surface with AI may help predict crop yield, forest health, and the effects of drought events benson2023forecasting (Fig. 3a, top). Similar DL drought detectors can be explained with xAI, identifying the most salient regions in time and space and attributing their main drivers dikshit2021interpretable (Fig. 3a, bottom).

4.2 Heatwaves: prediction, attribution and understanding

In the context of climate change, heat extremes are becoming more frequent and intense barriopedro_heat_2023 , but still regional trends are uncertain teng2022warming . For example, a notable warming hotspot in western Europe was identified with heat extremes increasing faster than predicted by the current state-of-the-art climate models vautard_heat_2023 . Understanding the physical drivers of heat extremes is crucial for accurate seasonal prediction and projection of these events. Several drivers of heat extremes remain uncertain though, specifically long-term changes in atmospheric circulation as well as land-atmosphere interactions, that may amplify or enhance specific events barriopedro_heat_2023 .

AI methods can accurately detect and predict heat extremes on different spatial and temporal scales, some lead times and aggregations using deep learning chattopadhyay_analog_2020 ; jacques2022deep ; fister_accurate_2023 , causal-informed ridge regression vijverberg2022role , or hybrid models van_straaten_CorrectingSubseasonal_2023 , among others. These approaches often require dimensionality reduction tools to select predictors from high-dimensional climate data (Fig. 3c). Expressive lower-dimensional feature representations have been extracted from high-dimensional data using VAEs, which can be used to improve the prediction of heat extremes miloshevich2023probabilistic . New methods, such as the EarthFormer, explore transformer networks specifically for the prediction of temperature anomalies, where encoder/decoder structures are combined with spatial attention layers gao2023earthformer . Next to predicting heatwaves, physical drivers of heatwaves can be inferred from the ML models as well. For example, XAI tools can be applied to discover physical drivers of heat extremes, e.g. van_straaten_CorrectingSubseasonal_2023 . Apart from XAI, causal inference methods have been used to understand the drivers of extreme temperatures vijverberg2022role .

Attribution studies for temperature and precipitation have increasingly used AI techniques barnes2019viewing . When it comes to heatwaves the identification of circulation-induced vs. thermodynamical changes becomes crucial. Several statistical and ML methods are currently in use for this purpose singh_circulation_2023 . The next vital step for attribution studies is to use such ML methods to understand the contribution of dynamical changes to the observed trends. This will lighten the discrepancies between circulation trends in climate model observations and reanalysis vautard_heat_2023 . Finally, recent development shows that ML techniques can be successfully applied for rare event modeling, sampling, and prediction miloshevich2023probabilistic .

4.3 Wildfires: predict, explain and understand

Most wildfire damage stems from a few extreme events. Predicting these events is vital for effective fire management and ecosystem conservation. Climate change is expected to increase the frequency, size, and severity of extreme wildfires, exacerbating fire weather conditions jones2022global . Traditional firefighting approaches are increasingly inadequate, with many extreme fires burning until naturally extinguished el_garroussi_europe_2024 . The complexity of nonlinear interactions among fire drivers across various scales hinders the predictability of fire behavior. Therefore, enhancing models to better understand and predict conditions leading to large, uncontrollable fires is crucial.

DL methods can associate weather predictions, satellite observations, and burned area datasets to model wildfires achieving better predictability than conventional methods. DL has been used successfully to forecast wildfire danger kondylatos_wildfire_2022 and for susceptibility map** zhang_forest_2019 . In the short term, wildfires are driven by the daily weather variations but also by the cumulative effects of vegetation and drought. On sub-seasonal to seasonal scales, where weather forecasts are less reliable, wildfires are modulated by Earth system large-scale processes, such as teleconnections. Early work has shown that DL models can leverage information from teleconnections to improve long-term wildfire forecasting li2023attentionfire_v1 . Apart from predicting wildfires, it is very important to understand the reasons behind the predictions to support decision-making and fire management. In this context, xAI can aid in identifying what drives wildfires, supporting, for example, the differentiation and management of various fire types, such as wind-driven and drought-driven fires kondylatos_wildfire_2022 .

Under global warming el_garroussi_europe_2024 , in the coming years, we expect an increase in the frequency of unstable atmospheric conditions that can cause Pyrocumulonimbus (pyroCbs), i.e., storm clouds that generate their weather fronts and can make wildfire behavior unpredictable fromm_understanding_2022 . Despite the risk posed by pyroCbs, the conditions leading to their occurrence and evolution are still poorly understood, and their causal mechanisms are uncertain. AI, in combination with causal inference, can advance the detection, forecasting, and understanding of the drivers of pyroCb events salas2022identifying .

4.4 Floods: from detection to communication of risk

Studying floods is crucial as they are the most frequent and costly natural disasters, affecting millions annually and causing over $40 billion in damages worldwide each year Jonkman2005Global . Develo** novel methods for flood detection enhances EWSs, reducing fatalities by up to 40%, while effective risk communication ensures better preparedness and response, potentially saving thousands of lives.

In July 2021, intense rainfall in Germany, Belgium, France, and the UK triggered severe flash floods, particularly devastating the Ahr River region in Germany, claiming nearly $200$ lives cornwall2021europe ; mohr2023multi . Alerts were issued by the European Flood Awareness System (EFAS) and disseminated through an international warning system. However, the floods exposed significant flaws in these systems (Fig. 3d). Damage to river monitoring infrastructure impeded data accuracy, and although meteorological forecasts predicted the rainfall two days in advance, flood predictions for smaller basins lacked precision and failed to consider debris flow and morphodynamics, leading to underestimations of flood severity.

AI offers promising avenues for enhancing flood management systems. Advanced global meteorological forecasting models powered by AI can rapidly process large data ensembles, providing more accurate probabilistic estimates even during extreme weather events. Furthermore, AI techniques such as ML-accelerated computational fluid dynamics could address the computational challenges in hydro-morpho-dynamic modeling, allowing for more precise predictions of stream flow and flood levels, particularly in ungauged basins kochkov2021machine .

Additionally, AI can aid in calibrating non-contact video gauges, potentially more robust than traditional methods, as they are not directly affected by the water. AI can also guide forensic analysis to evaluate exposure and vulnerability, employing multi-modal approaches to refine geospatial models at a local scale. However, the effectiveness of these AI-driven models depends on detailed knowledge of the local terrain, including potential bottlenecks like bridges and channels, and accurate data about societal vulnerability to floods.

Yet, AI can also transform how warnings are issued to improve communication and response strategies. For instance, AI-generated maps and photorealistic visualizations based on digital elevation models can depict expected inundation areas and damages lutjens2023physicallyconsistent . Moreover, AI can generate easily understandable, language-based warnings, both written and auditory, tailored for diverse populations, including the visually impaired. An LLM-based chatbot feature could enhance interactivity, providing real-time, personalized responses to emergency inquiries (see examples in Fig. 3d).

5 Conclusions and Outlook

This review highlighted the significant potential of AI for analyzing and modeling extreme events while also detailing the main difficulties and prospects associated with this emerging field. Integrating AI into extreme event analysis faces several challenges, including data management issues like handling dynamic datasets, biases, and high dimensionality that complicate feature extraction. AI models also struggle with unclear statistical definitions of what is ‘extreme’. Furthermore, integrating AI with physical models poses substantial challenges yet offers promising opportunities for enhancing model accuracy and reliability. Trustworthiness concerns arise from the complexity and interpretability of ML models, the difficulty of generalizing across different contexts, and the quantification of uncertainty. Operational challenges include the complexity of AI outputs which hinder interpretation by non-experts, resistance to AI adoption due to concerns over reliability and fairness, and the need for frameworks that facilitate transparent and ethical integration of AI insights into decision-making processes.

The previous challenges compromise reproducibility and comparability of ML models in analyzing extreme events. These challenges are exacerbated by data scarcity, lack of transparency in model configurations, and the use of proprietary tools. Additionally, interdisciplinary differences hinder consistency and comparability. Effective solutions demand robust, transparent methodologies, inclusive data-sharing practices, and frameworks that support cross-disciplinary collaboration.

We overviewed and emphasized the importance of develo** operational, explainable, and trustworthy AI systems. Addressing these challenges requires a coordinated effort across disciplines involving AI researchers, environmental and climate scientists, field experts, and policymakers. This collaborative approach is crucial for advancing AI applications in extreme event analysis and ensuring that these technologies are adapted to real-world needs and constraints. From an operational perspective, adapting AI solutions to real-time data integration, model deployment, and resource allocation highlights the need for systems that can function within the operational frameworks of disaster management and risk mitigation. Besides, methodological improvements are still needed in model evaluation and benchmarking to alleviate issues like overfitting and enhance AI systems’ generalization capabilities.

Looking forward, there are significant areas for further exploration and improvement. These include the development of benchmarks specific to extreme events, enhanced integration of domain knowledge to improve data fusion and model training, and the creation of robust, scalable AI systems capable of adapting to the dynamic nature of extreme events. Recent LLMs harvest vast amounts of domain knowledge embedded in the literature, promising significant advances in communicating risk.

As we advance, the ultimate goal is to harness AI’s potential to substantially benefit society, particularly by enhancing our capacity to manage and respond to extreme events. Through dedicated research and collaborative innovation, AI can become a cornerstone in our strategy to understand and mitigate the impacts of these challenging and elusive phenomena.