Extreme hydro-meteorological events, such as floods, have become increasingly prevalent in recent years, posing significant threats to communities worldwide. Effective forecasting and warning systems are essential for mitigating impacts of these natural disasters and for protecting both properties and human lives. In particular, flood risk management in populated and orographically complex environ ments presents critical challenges due to increasing climatic extremes, urbanization, and the limitations of traditional modeling approaches. This thesis addresses these challenges through a trilogy of complementary studies focused on improving real-time flood fore casting and early warning capabilities in the Lombardy Region of northern Italy, with particular attention to the "hydraulic node of Milan". The first study develops and validates a methodology for defining empirical, catchment specific empirical Rainfall Thresholds (RTs) for flood warning, integrating soil moisture conditions through a novel "equivalent rainfall" approach. Based on 25 years of hydrom eteorological data, this work demonstrates that RTs incorporating antecedent moisture information significantly reduce false alarms and can be easily adopted by civil protec tion agencies, offering a lightweight yet robust alternative or supplement to model-based systems, as well as providing a standardized procedure to create rainfall thresholds at the “homogeneous zone” scale, which is the spatial unit used by the Italian national civil protection system for issuing color-coded warnings. The second study tackles the issue of spatial uncertainty in Quantitative Precipitation Forecasts (QPF) from convection-permitting models. By analyzing historical convective events, it identifies systematic spatial biases and proposes a kernel-based ensemble shifting technique that enables to quantify forecast spatial uncertainty through the generation of ensemble forecasts. This method enhances the reliability of flood predictions under uncertain meteorological forcing, supporting probabilistic early warning strategies. The third study benchmarks classical and deep learning time-series models, including lin ear Autoregressive Moving-Average with Exogenous Inputs (ARMAX), Non-linear AR MAX(NARX), Long-Short Term (LSTM) and Gated Recurrent Unit (GRU) neural net works, for flood prediction using only rainfall and hydrometric level time series. It shows that recurrent neural networks, even in simplified configurations, outperform traditional models in terms of predictive accuracy and data efficiency. Sensitivity analyses were con ducted to further quantify the minimum data requirements for reliable deployment in operational flood forecasting systems. The approaches combine machine and deep learning, empirical hydrology, and meteo rological forecast uncertainty quantification, offering practical tools for civil protection authorities to improve their flood warning systems. Together, these contributions could establish an integrated framework for operational flood forecasting that is adaptive, data efficient, and applicable at both the localized and regional scale
Gli eventi idro-meteorologici estremi, come le alluvioni, sono diventati più frequenti negli ultimi anni, rappresentando gravi minacce per le comunità a livello globale. Sistemi efficaci di previsione e allertamento sono essenziali per mitigare gli impatti di questi disastri naturali e per proteggere sia i beni materiali che le vite umane. In particolare, la gestione del rischio alluvioni in ambienti popolati e orograficamente complessi presenta sfide critiche a causa dell’accentuarsi degli estremi climatici, dell’urbanizzazione crescente e delle limitazioni degli approcci modellistici tradizionali. La presente tesi affronta tali sfide attraverso una trilogia di studi complementari, finalizzati al miglioramento delle capacità di previsione e allerta precoce in tempo reale per le alluvioni nella Regione Lombardia, con particolare attenzione al “nodo idraulico di Milano”. Il primo studio sviluppa e valida una metodologia per la definizione di Soglie di Pioggia (Rainfall Thresholds, RT) empiriche e specifiche per bacino, integrando le condizioni di umidità del suolo attraverso un innovativo approccio basato sulla “pioggia equivalente”. Basandosi su 25 anni di dati idro-meteorologici, questo lavoro dimostra che le soglie che incorporano l’informazione sull’umidità antecedente riducono significativamente i falsi allarmi e possono essere facilmente adottate dalle agenzie di protezione civile, offrendo un’alternativa leggera ma robusta – o un’integrazione – ai sistemi basati su modelli. Inoltre, fornisce una procedura standardizzata per la creazione di soglie pluviometriche alla scala di “zona omogenea”, che rappresenta l’unità spaziale utilizzata dal sistema nazionale di allertamento della protezione civile per l’emissione degli avvisi codificati a colori. Il secondo studio affronta il problema dell’incertezza spaziale nelle Previsioni Quantitative di Precipitazione (QPF) generate da modelli numerici a risoluzione convettiva. Analizzando eventi convettivi storici, individua bias spaziali sistematici e propone una tecnica di traslazione probabilistica dell’insieme (ensemble shifting) basata su kernel, che consente di quantificare l’incertezza spaziale nelle previsioni attraverso la generazione di previsioni ensemble. Questo metodo migliora l’affidabilità delle previsioni di piena in condizioni di forzante meteorologica incerta, supportando strategie di allerta precoce di tipo probabilistico. Il terzo studio confronta modelli classici e di apprendimento profondo per serie temporali, tra cui il modello lineare Autoregressivo a Media Mobile con Input Esogeni (ARMAX), il modello non lineare ARMAX (NARX), nonché reti neurali ricorrenti Long-Short Term Memory (LSTM) e Gated Recurrent Unit (GRU), per la previsione delle piene utilizzando unicamente serie temporali di pioggia e livelli idrometrici. Lo studio dimostra che le reti neurali ricorrenti, anche in configurazioni semplificate, superano i modelli tradizionali in termini di accuratezza predittiva ed efficienza nell’uso dei dati. Sono state inoltre condotte analisi di sensitività per quantificare i requisiti minimi di dati necessari a un impiego affidabile nei sistemi operativi di previsione delle piene. Gli approcci proposti combinano apprendimento automatico e profondo, idrologia empirica e quantificazione dell’incertezza delle previsioni meteorologiche, offrendo strumenti pratici alle autorità di protezione civile per il miglioramento dei sistemi di allerta in caso di alluvione. Nel loro insieme, questi contributi potrebbero costituire un quadro integrato per la previsione operativa delle piene, adattivo, efficiente nell’uso dei dati e applicabile sia su scala locale che regionale.
Data-driven flood forecasting for civil protection: integrating machine learning, empirical rainfall thresholds and forecast uncertainty
Enrico, Gambini
2025
Abstract
Extreme hydro-meteorological events, such as floods, have become increasingly prevalent in recent years, posing significant threats to communities worldwide. Effective forecasting and warning systems are essential for mitigating impacts of these natural disasters and for protecting both properties and human lives. In particular, flood risk management in populated and orographically complex environ ments presents critical challenges due to increasing climatic extremes, urbanization, and the limitations of traditional modeling approaches. This thesis addresses these challenges through a trilogy of complementary studies focused on improving real-time flood fore casting and early warning capabilities in the Lombardy Region of northern Italy, with particular attention to the "hydraulic node of Milan". The first study develops and validates a methodology for defining empirical, catchment specific empirical Rainfall Thresholds (RTs) for flood warning, integrating soil moisture conditions through a novel "equivalent rainfall" approach. Based on 25 years of hydrom eteorological data, this work demonstrates that RTs incorporating antecedent moisture information significantly reduce false alarms and can be easily adopted by civil protec tion agencies, offering a lightweight yet robust alternative or supplement to model-based systems, as well as providing a standardized procedure to create rainfall thresholds at the “homogeneous zone” scale, which is the spatial unit used by the Italian national civil protection system for issuing color-coded warnings. The second study tackles the issue of spatial uncertainty in Quantitative Precipitation Forecasts (QPF) from convection-permitting models. By analyzing historical convective events, it identifies systematic spatial biases and proposes a kernel-based ensemble shifting technique that enables to quantify forecast spatial uncertainty through the generation of ensemble forecasts. This method enhances the reliability of flood predictions under uncertain meteorological forcing, supporting probabilistic early warning strategies. The third study benchmarks classical and deep learning time-series models, including lin ear Autoregressive Moving-Average with Exogenous Inputs (ARMAX), Non-linear AR MAX(NARX), Long-Short Term (LSTM) and Gated Recurrent Unit (GRU) neural net works, for flood prediction using only rainfall and hydrometric level time series. It shows that recurrent neural networks, even in simplified configurations, outperform traditional models in terms of predictive accuracy and data efficiency. Sensitivity analyses were con ducted to further quantify the minimum data requirements for reliable deployment in operational flood forecasting systems. The approaches combine machine and deep learning, empirical hydrology, and meteo rological forecast uncertainty quantification, offering practical tools for civil protection authorities to improve their flood warning systems. Together, these contributions could establish an integrated framework for operational flood forecasting that is adaptive, data efficient, and applicable at both the localized and regional scale| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_Enrico_Gambini___PhD__A4.pdf
embargo fino al 04/07/2028
Licenza:
Tutti i diritti riservati
Dimensione
60.08 MB
Formato
Adobe PDF
|
60.08 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/356162
URN:NBN:IT:POLIMI-356162