Flood prediction is crucial for risk management and adaptation to reduce the impacts caused by increasing hydrological risk. Current continental or global hydrological models, such as the Global Flood Awareness System (GloFAS, part of the Copernicus Emergency Management Service), are used to inform early warning and large-scale adaptation actions, especially for transboundary river basins. These models also support humanitarian organizations, national and international bodies, complementing the few tools available in some developing regions and aiding decisions and actions to mitigate the impacts of extreme events. However, these models often suffer from limitations due to poor accuracy and large predictive errors compared to observations. This thesis explores the application of a deep learning model, based on a Long Short-Term Memory (LSTM) architecture, to improve the historical simulation (reanalysis) and forecasts of GloFAS. The effectiveness and value of these improvements are demonstrated for early warning and decision-making in the Zambezi basin in southern Africa, which is exposed to frequent tropical cyclones and consequent floods. A comparative analysis with discharge observations revealed a significant bias and poor predictive accuracy in both the reanalysis and forecasts of the GloFAS model in the Zambezi region. Using a deep learning-based post-processing methodology, this bias was substantially reduced and accuracy improved, leading to enhanced predictive performance, as demonstrated by various traditional metrics such as Root Mean Square Error (RMSE) and Kling-Gupta Efficiency (KGE), and end-user-oriented metrics such as False Alarm Ratios (FAR) and Probability of Detection (POD). Furthermore, this study validated the improvements made to GloFAS flood hazard maps (available at increasingly high resolution) with a new dataset of observed floods from satellite (Landsat), validated globally with impact data (EM-DAT). To construct this dataset, a state-of-the-art image segmentation method for flood detection, called Otsu was built and consequently compared with an alternative algorithm based on the Modified Normalized Difference Water Index (MNDWI). Using an impact dataset (EM-DAT) as a reference, the results indicated that the Otsu algorithm significantly outperformed the MNDWI alternative, providing a more accurate basis for evaluating GloFAS maps of simulated and forecasted floods. The integration of deep learning techniques, satellite imagery, and advanced algorithms to evaluate and improve flood forecasting represents a significant advancement in hydrological prediction, offering practical solutions to enhance forecast accuracy and promoting better flood management and risk reduction strategies. Future research could extend these methodologies to other regions and models, further contributing to global efforts to increase flood resilience.
La previsione delle inondazioni è fondamentale per una efficace gestione e adattamento al rischio, al fine di ridurre gli impatti causati da un rischio idrologico crescente. Gli attuali modelli idrologici continentali o globali, come il Global Flood Awareness System (GloFAS, del Copernicus Emergency Management Service), sono usati per informare l'allerta precoce e azioni di adattamento a larga scala, specialmente per bacini idrografici transfrontalieri. Tali modelli inoltre supportano organizzazioni umanitarie, enti nazionali e internazionali, complementando i pochi strumenti disponibili in alcune regioni in via di sviluppo e supportando decisioni e azioni per ridurre gli impatti di eventi estremi. Tuttavia, questi modelli spesso soffrono di limitazioni date dalla scarsa accuratezza e grandi errori predittivi rispetto alle osservazioni. Questa tesi esplora l'applicazione di un modello di deep learning, basato su una architettura di tipo Long Short-Term Memory (LSTM), per migliorare la simulazione storica (rianalisi) e le previsioni di GloFAS. L’efficacia ed il valore di questi miglioramenti sono dimostrati per l'allerta precoce e la presa di decisioni per il bacino dello Zambezi in Africa meridionale, che è esposto a un frequente rischio di cicloni tropicali e conseguenti inondazioni. Un'analisi comparativa con osservazioni di portate ha rivelato un notevole bias e una scarsa accuratezza predittiva sia nella rianalisi che nelle previsioni del modello GloFAS nella regione dello Zambezi. Utilizzando una metodologia di post-elaborazione basata sul deep learning, questo bias è stato sostanzialmente ridotto e l’accuratezza migliorata, portando a prestazioni predittive migliorate, come dimostrato da diverse metriche tradizionali, come Root Mean Square Error (RMSE) e Kling-Gupta Efficiency (KGE), e da metriche orientate agli utenti finali, come False Alarm Ratios (FAR) e Probability of Detection (POD). Inoltre, questo studio ha validato i miglioramenti apportati a GloFAS, specialmente alle sue mappe di inondazione a più alta risoluzione, con un nuovo dataset di inondazioni osservate da satellite (Landsat), validato a scala globale con dati di impatti (EM-DAT). Per costruire questo dataset, un metodo di stato dell’arte di segmentazione delle immagini per il rilevamento delle inondazioni, chiamato Otsu, è stato confrontato con un algoritmo alternativo, basato sull’indice Modified Normalized Difference Water Index (MNDWI). Usando un dataset di impatti (EM-DAT) come riferimento, i risultati hanno indicato che l’algoritmo Otsu ha una performance significativamente migliore dell’alternativa con MNDWI, fornendo una base più accurata per valutare le mappe di GloFAS delle inondazioni simulate e previste. L’integrazione di tecniche di deep learning, immagini da satellite e algoritmi avanzati per valutare e migliorare la previsione delle inondazioni rappresenta un progresso significativo nella previsione idrologica, offrendo soluzioni pratiche per migliorare l’accuratezza delle previsioni e favorendo una migliore gestione delle inondazioni e strategie di adattamento e riduzione del rischio. Ricerche future potrebbero estendere queste metodologie ad altre regioni e modelli, contribuendo ulteriormente agli sforzi globali di aumentare la resilienza alle inondazioni.
Enhancing global flood detection and forecasting using deep learning
MIR, MOHID FAYAZ
2023/2024
Abstract
Flood prediction is crucial for risk management and adaptation to reduce the impacts caused by increasing hydrological risk. Current continental or global hydrological models, such as the Global Flood Awareness System (GloFAS, part of the Copernicus Emergency Management Service), are used to inform early warning and large-scale adaptation actions, especially for transboundary river basins. These models also support humanitarian organizations, national and international bodies, complementing the few tools available in some developing regions and aiding decisions and actions to mitigate the impacts of extreme events. However, these models often suffer from limitations due to poor accuracy and large predictive errors compared to observations. This thesis explores the application of a deep learning model, based on a Long Short-Term Memory (LSTM) architecture, to improve the historical simulation (reanalysis) and forecasts of GloFAS. The effectiveness and value of these improvements are demonstrated for early warning and decision-making in the Zambezi basin in southern Africa, which is exposed to frequent tropical cyclones and consequent floods. A comparative analysis with discharge observations revealed a significant bias and poor predictive accuracy in both the reanalysis and forecasts of the GloFAS model in the Zambezi region. Using a deep learning-based post-processing methodology, this bias was substantially reduced and accuracy improved, leading to enhanced predictive performance, as demonstrated by various traditional metrics such as Root Mean Square Error (RMSE) and Kling-Gupta Efficiency (KGE), and end-user-oriented metrics such as False Alarm Ratios (FAR) and Probability of Detection (POD). Furthermore, this study validated the improvements made to GloFAS flood hazard maps (available at increasingly high resolution) with a new dataset of observed floods from satellite (Landsat), validated globally with impact data (EM-DAT). To construct this dataset, a state-of-the-art image segmentation method for flood detection, called Otsu was built and consequently compared with an alternative algorithm based on the Modified Normalized Difference Water Index (MNDWI). Using an impact dataset (EM-DAT) as a reference, the results indicated that the Otsu algorithm significantly outperformed the MNDWI alternative, providing a more accurate basis for evaluating GloFAS maps of simulated and forecasted floods. The integration of deep learning techniques, satellite imagery, and advanced algorithms to evaluate and improve flood forecasting represents a significant advancement in hydrological prediction, offering practical solutions to enhance forecast accuracy and promoting better flood management and risk reduction strategies. Future research could extend these methodologies to other regions and models, further contributing to global efforts to increase flood resilience.File | Dimensione | Formato | |
---|---|---|---|
MasterThesis_Mohid_Fayaz_Mir.pdf
accessibile in internet per tutti
Descrizione: Enhancing Global Flood detection and Forecasting using Deep Learning
Dimensione
26.22 MB
Formato
Adobe PDF
|
26.22 MB | 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/10589/226839