Table 1 Comprehensive categorization of extreme events analysis
From: Artificial intelligence for modeling and understanding extreme weather and climate events
Aspect | Task | Standard methods | Current AI methods | |
|---|---|---|---|---|
Data | Data collection, harmonization, preprocessing | Gather, standardize, and clean relevant multi-source data for the down-stream task | Manual normalization and detrending; statistical methods for denoising, gap-filling, and data fusion | Automated ML/DL-based pipelines for denoising, gap-filling, and data fusion; anomaly detection |
Feature selection and extraction | Find informative low-dimensional representations of the data | Statistical tests and dimensionality reduction, manual feature selection, and feature engineering | Automated feature selection via feature importance; deep feature extraction and embedded methods; causal feature selection and representation learning | |
Simulation of extremes | Simulate extremes under varying scenarios with correct frequency and amplitude | Physical models, Monte Carlo simulations, extreme value theory | ML-based, physics-informed, and hybrid generative models | |
Modeling | Detection | Detect and identify extreme events geographically and over time | Threshold- or percentile-based methods, extreme value theory | Outlier detection, one-class classification, reconstruction-based models, probabilistic methods |
Prediction | Make predictions on potential future extreme events | Numerical weather prediction models, expert judgment, heuristics, time-series analysis | Deterministic and probabilistic ML/DL forecasting models, recurrent neural networks | |
Impact assessment | Estimate the effects of extreme events on society, the economy, and the natural environment | Expert judgment, heuristics, field data collection, statistical impact analysis | Simulation-based ML/DL models, deterministic and probabilistic ML/DL methods for predictive impact modeling | |
Understanding and trustworthiness | Explainability | Make the rationale of a black-box ML method understandable by humans | Transparent and interpretable by design models (e.g., linear approximations, decision trees) | Distillation methods, feature attribution methods, attention mechanisms |
Causality | Identify and analyze factors directly leading to the occurrence of extreme events | Statistical and correlation-based analysis, simulation-based, and experimental studies | Causal inference, counterfactual explanations | |
Extreme event attribution | Quantify influence of anthropogenic forcings on extreme events | Statistical event attribution, storyline approaches, expert judgment, and elicitation | Weather and climate emulation models, ensemble techniques | |
Uncertainty quantification | Measure the reliability of predictions, identify and quantify sources of uncertainty | Confidence intervals, sensitivity analysis, parametric distributions | Bayesian methods, ensemble methods | |
Last mile | Operationalization | Integrate trustworthy predictive models and real-time data analysis into early warning systems | Manual integration into decision systems | Automated integration using APIs and ML-driven decision support, cloud-based deployment |
Communication of risk and ethical aspects | Convey information on extreme risks, ensuring ethical transparency, community preparedness and response | Adherence to ethical guidelines and standards, risk reports | Ethical and fairness-aware AI models, causality for quantification of biases | |
Policy and Decision-making | Manage and mitigate impacts of extreme events, ensuring resilience and safety | Risk assessment and management, cost-benefit analysis, multi-criteria decision analysis, adaptive management, and community-based participatory approaches | Time series analysis, ensemble methods, anomaly detection, clustering, reinforcement learning |
