Tropical cyclones are intense atmospheric systems that pose significant threats to life, ecosystems, and infrastructures worldwide. Tropical Cyclogenesis (TCG) occurs when a tropical disturbance intensifies into a fully-formed cyclone. The difficulty of the prediction of TCG is due to the complex interplay of environmental variables and the limited number of data available. Large-scale indices like Genesis Potential Indices (GPIs) are commonly used to estimate the likelihood of TCG. However, their predictive skill is often limited by regional discrepancies and computational complexity. This thesis proposes a novel approach to improve TCG prediction by leveraging causal discovery and machine learning techniques. The methodology identifies key causal relationships between various atmospheric and oceanic variables and their role in TCG prediction in both pixel-wise and basin-wide analyses. High-resolution observational data and causal discovery algorithms are used to uncover the underlying drivers of TCG. The results highlight significant regional variations in causal factors and emphasize the importance of localized environmental conditions. The performance of the proposed machine learning-based prediction models is compared with the traditional Emanuel-Nolan GPI index, demonstrating superior accuracy in predicting TCG in some ocean basins, such as North West Pacific. The key contribution of this thesis is the identification of critical causal factors for TCG, developing a novel machine learning framework for cyclone prediction to validate the efficacy of the selected causal relevant variables. These findings suggest that machine learning models can improve TCG forecasting accuracy, with promising results that could enhance difficult tasks such as disaster risk management and future climate predictions. Ultimately, this research represents a step toward more reliable, region-specific forecasting of tropical cyclones, offering potential benefits for mitigation and adaptation strategies in cyclone-prone regions.
I cicloni tropicali sono sistemi atmosferici intensi che rappresentano minacce significative per la vita, gli ecosistemi e le infrastrutture a livello globale. La genesi del ciclone tropicale (TCG) si verifica quando una perturbazione tropicale si intensifica trasformandosi in un ciclone completamente formato. La difficoltà nella previsione della TCG è dovuta alla complessa interazione tra le variabili ambientali e alla quantità limitata di dati disponibili. Indici su larga scala come i Genesis Potential Indices (GPIs) sono comunemente utilizzati per stimare la probabilità di TCG. Tuttavia, la loro capacità predittiva è spesso limitata da discrepanze regionali e dalla complessità computazionale. Questa tesi propone un approccio innovativo per migliorare la previsione del TCG, sfruttando tecniche di causal discovery e di machine learning. La metodologia identifica le principali relazioni causali tra varie variabili atmosferiche e oceaniche e il loro ruolo nella previsione del TCG, sia in analisi pixel-per-pixel che su scala di bacino. Vengono utilizzati dati osservativi ad alta risoluzione e algoritmi di causal discovery per svelare i fattori sottostanti del TCG. I risultati evidenziano significative variazioni regionali nei fattori causali e sottolineano l'importanza delle condizioni ambientali locali. Le prestazioni dei modelli di previsione basati su machine learning proposti vengono confrontate con l'indice tradizionale Emanuel-Nolan GPI, mostrando una maggiore precisione nel prevedere il TCG in alcuni bacini oceanici, come il Pacifico del Nord-Ovest. Il contributo principale di questa tesi è l'identificazione dei fattori causali critici per la TCG, attraverso lo sviluppo di un nuovo framework di machine learning per la previsione dei cicloni, finalizzato a validare l'efficacia delle variabili causali selezionate. Questi risultati suggeriscono che i modelli di machine learning possono migliorare la precisione delle previsioni della TCG, offrendo prospettive promettenti per affrontare compiti complessi come la gestione del rischio di catastrofi e le previsioni climatiche future. In ultima analisi, questa ricerca rappresenta un passo verso previsioni più affidabili e specifiche per regione dei cicloni tropicali, offrendo potenziali benefici per strategie di mitigazione e adattamento nelle regioni vulnerabili ai cicloni.
Investigating tropical cyclogenesis drivers with causal analysis and machine learning
ARNEODO, ALICE
2023/2024
Abstract
Tropical cyclones are intense atmospheric systems that pose significant threats to life, ecosystems, and infrastructures worldwide. Tropical Cyclogenesis (TCG) occurs when a tropical disturbance intensifies into a fully-formed cyclone. The difficulty of the prediction of TCG is due to the complex interplay of environmental variables and the limited number of data available. Large-scale indices like Genesis Potential Indices (GPIs) are commonly used to estimate the likelihood of TCG. However, their predictive skill is often limited by regional discrepancies and computational complexity. This thesis proposes a novel approach to improve TCG prediction by leveraging causal discovery and machine learning techniques. The methodology identifies key causal relationships between various atmospheric and oceanic variables and their role in TCG prediction in both pixel-wise and basin-wide analyses. High-resolution observational data and causal discovery algorithms are used to uncover the underlying drivers of TCG. The results highlight significant regional variations in causal factors and emphasize the importance of localized environmental conditions. The performance of the proposed machine learning-based prediction models is compared with the traditional Emanuel-Nolan GPI index, demonstrating superior accuracy in predicting TCG in some ocean basins, such as North West Pacific. The key contribution of this thesis is the identification of critical causal factors for TCG, developing a novel machine learning framework for cyclone prediction to validate the efficacy of the selected causal relevant variables. These findings suggest that machine learning models can improve TCG forecasting accuracy, with promising results that could enhance difficult tasks such as disaster risk management and future climate predictions. Ultimately, this research represents a step toward more reliable, region-specific forecasting of tropical cyclones, offering potential benefits for mitigation and adaptation strategies in cyclone-prone regions.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Arneodo_Tesi_01.pdf
accessibile in internet per tutti
Descrizione: Testo tesi
Dimensione
5.2 MB
Formato
Adobe PDF
|
5.2 MB | Adobe PDF | Visualizza/Apri |
2024_12_Arneodo_Executive_Summary_02.pdf
accessibile in internet per tutti
Descrizione: Testo executive summary
Dimensione
514.13 kB
Formato
Adobe PDF
|
514.13 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/231077