Detection and Prediction of Extreme Weather Events using Machine Learning and Deep Learning Techniques

• Autores: Jorge Peréz Aracil
• Directores de la Tesis: Pedro Antonio Gutiérrez Peña (dir. tes.) , Sancho Salcedo Sanz (dir. tes.)
• Lectura: En la Universidad de Córdoba (ESP) ( España ) en 2026
• Idioma: inglés
• Tribunal Calificador de la Tesis: José Cristobal Riquelme Santos (presid.) , Juan Carlos Fernández Caballero (secret.) , Silvia Jiménez Fernández (voc.)
• Enlaces
  • Tesis en acceso abierto en: Helvia
• Resumen

  • Extreme weather events, particularly heatwaves (HWs) and extreme temperatures, represent one of the most pressing challenges of contemporary climate change, with devastating impacts on human health, agricultural systems, energy infrastructures, and natural ecosystems. This thesis presents a comprehensive framework of Artificial Intelligence (AI)-based methodologies to enhance the understanding, detection, prediction, and attribution of extreme climate events, with special emphasis on extreme temperature phenomena.

    This thesis addresses three fundamental objectives that constitute the main contributions of this work. First, an explainable and usable framework is developed to identify key drivers of extreme climate events through the STCO-FS (Spatio-temporal Cluster-Optimised Feature Selection) methodology. This innovative approach combines spatial clustering techniques with metaheuristic optimisation to identify relevant variables, considering both spatial and temporal dimensions, thereby overcoming the limitations of traditional feature selection methods that fail to capture the spatio-temporal dependencies inherent in climate data adequately.

    Second, a Deep Learning (DL)-based methodology is proposed to enhance the reconstruction of extreme climate events by developing the Autoencoder (AE)-Analogue Method (AM). This hybrid approach combines the capability of AEs to capture complex non-linear patterns in lower-dimensional latent spaces with the statistical robustness of traditional analogue methods. Results demonstrate significant improvements in the probabilistic reconstruction of historical European heatwaves, including emblematic events such as the 2003 HW in France, the 1995 HW in Spain, and the 2010 heatwave in Russia, providing more accurate and realistic temperature distributions than conventional analogue methods.

    Third, advanced DL architectures are developed for long-term prediction of extreme temperatures, including two innovative approaches: AE+Multi-layer Perceptron (MLP) and AE+AE. These hybrid architectures leverage the capability of autoencoders to extract meaningful representations from complex meteorological fields, combining them with specialised neural networks for temporal prediction. Experiments conducted across multiple European locations demonstrate the superiority of these methods over traditional approaches such as persistence and climatology, especially during heatwave periods where predictive accuracy is crucial for early warning systems.

    The obtained results demonstrate significant advances in multiple aspects of extreme event analysis. The STCO-FS framework successfully identifies key drivers with high precision, achieving F1-scores exceeding 0.85 in heatwave detection. The AE-AM method exhibits substantial improvements in extreme event reconstruction, with temperature distributions that are closer to real observations compared to traditional analogue methods. Long-term prediction architectures demonstrate significantly lower prediction errors, particularly during extreme events where traditional methods tend to fail.

    This thesis makes a significant contribution to the state of the art in multiple dimensions. From a methodological perspective, it introduces innovative techniques that effectively combine traditional machine learning with DL, leveraging the strengths of both approaches. From an applied perspective, it provides practical tools for climate risk management, early warning systems, and studies on the attribution of extreme events. From a scientific standpoint, it enhances understanding of the physical mechanisms underlying extreme events through the use of explainable AI techniques.

    The implications of this work extend beyond the academic realm, offering direct applications in operational meteorological services, emergency management, urban planning, and climate change adaptation. The developed methods can be integrated into existing numerical weather prediction systems to improve the representation of extreme events, contributing to societal resilience against the growing challenges of climate change. Actually, the United Nations, through the Focus Group on AI for Natural Disaster Management (FG-AI4NDM), has considered the case study developed for the Ada River Basin, shown in this thesis, of general interest. It is being considered for further analysis, and the methodology will be evaluated for its implementation to aid in driver identification of climate extreme events.

    The thesis demonstrates that AI-based approaches, when properly designed and validated, can significantly enhance our capability to understand, predict, and respond to extreme climate events. The methodologies developed