Detection and attribution of heat waves with the Multivariate Autoencoder Flow-Analogue Method (MvAE-AM)

Heat waves (HWs) are complex, multivariate, extreme weather events that cause significant harm to human health, ecosystems, and economies. Correct detection and attribution of HWs to anthropogenic climate change is important to better understand the underlying mechanisms and to improve predictions....

Descripción completa

Detalles Bibliográficos
Autores: Marina, Cosmin Madalin|||0000-0002-5849-6673, Pérez Aracil, Jorge|||0000-0002-4456-9886, McAdam, Ronan, Lorente Ramos, Eugenio, Luther, Niklas, Zorita, Eduardo, Scoccimarro, Enrico, Luterbacher, Jürg, Xoplaki, Elena, Salcedo Sanz, Sancho|||0000-0002-4048-1676
Tipo de recurso: artículo
Fecha de publicación:2026
País:España
Institución:Universidad de Alcalá (UAH)
Repositorio:e_Buah Biblioteca Digital Universidad de Alcalá
Idioma:inglés
OAI Identifier:oai:ebuah.uah.es:10017/67021
Acceso en línea:http://hdl.handle.net/10017/67021
https://dx.doi.org/10.1016/j.atmosres.2025.108409
Access Level:acceso abierto
Palabra clave:Heat waves
Attribution
Analogue method
Autoencoders
Deep learning
Explainable AI
Telecomunicaciones
Telecommunication
Descripción
Sumario:Heat waves (HWs) are complex, multivariate, extreme weather events that cause significant harm to human health, ecosystems, and economies. Correct detection and attribution of HWs to anthropogenic climate change is important to better understand the underlying mechanisms and to improve predictions. In this work, we address this issue and propose a multivariate version of a hybrid approach to reconstruct heat waves, consisting of the AM and deep Autoencoders (MvEA-AM algorithm), improving existing less effective methods used until now, such as the multivariate Analogue Method (MvAM). The proposed hybrid approach produces a more reliable representation of the event than the classical MvAM for reconstructing and attributing HWs in Europe. The explainable and interpretable analysis of the obtained results is based on leveraging the SHapley Additive exPlanations (SHAP) method to explain deep learning algorithms, a capability that is not achievable with the MvAM. This explainability analysis shows that our model learns useful features during the training of the algorithm, which are aligned with the Physics of the problem, and employs the correct features during reconstruction and attribution analysis of the HWs considered.