Droughts can have a serious impact on health, agriculture, economy, energy and environment. Monitoring and forecasting such phenomena is an essential task for mitigation, adaptation, and reduction of future risks by policymakers, mainly for areas that are characterized by high water demand with respect to the available fresh water. The Nile River Basin is a region vulnerable to climate change facing high water stress, since almost half of the basin is an arid area, where the population relies on 93$%$ of their fresh water from the Nile River. Considering that the Nile waters are fully exploited to fulfill the high water demands downstream, climate change will exacerbate the water-related challenges, leading to an increased risk of water scarcity and food insecurity. This raises the necessity of a reliable drought detection index to support water managers and decision makers in avoiding such risks. Thus, this thesis focuses on the design of a basin-specific drought index over the Nile River Basin, which consists in an interesting case study characterized by climate and topographical heterogeneity. Also, the Nile transboundary waters are cause of conflicts among the sharing countries and still generate international discussion. The index design is performed through the application of a machine learning FRamework for Index based Drought Analysis: FRIDA cite{zaniolo2018automatic}. The proposed framework was applied on the sub-basin scale, due to the complexity and heterogeneity of the system. Thus, the Nile River Basin is divided into 10 sub-basins according to hydrology, climate, and land cover. The framework is based on 3 main steps: information retrievement, feature extraction, and index construction. The first is performed by the definition of a target variable and collection of prediction variables: the Normalized Drought Vegetation Index is chosen as the target variable, since the economy of the basin countries is highly dependent on agriculture, highlighting the ability of the target variable in detecting agricultural and meteorological droughts. The prediction variables are chosen based on literature, identifying the high correlation with drought events, long term effect, and case study characteristics. The second step is performed by using Wrapper for Quasi-Equally Informative Subset Selection as the tool for feature extraction according to the relatively better performance of wrappers than filters. The third step---index construction---performed using two alternative regression models (i.e., linear and Artificial Neural Networks). A common thread highlighted from the output results is the consistency of the study area basin subdivision criteria, however an additional criteria of considering the spatial resolution of the input variables can be an improvement, since the sub-basins have different surface areas, topographic, and climatic characteristics. Thus, accounting for noise filtering, avoiding the effect of extremes and outliers in a portion of the sub-basin. Furthermore, the results convey high performance of the designed basin specific drought indexes in reproducing the target variable, where the index predictive accuracy in 90$%$ of the sub-basins is outstanding; while the low performance in some cases is due to multiple factors, mainly land cover and surface area. The study contributes a further proof of the expected feature extraction efficiency in meeting the objectives of the proposed framework, reproducing a drought index for the Nile River Basin highly accurate in most of the cases. The application of FRIDA can provide an appropriate drought monitoring index that can adapt to basins with different characteristics.
La siccità può avere impatti molto seri sulla salute, l’agricoltura, l’economia, l’energia e l’ambiente. Monitorare e prevedere tale fenomeno e compito essenziale della politica per mitigare, adattare e ridurre rischi futuri, soprattutto per le aree caratterizzate da un’alta domanda di acqua rispetto alla disponibilità di acqua dolce. Il bacino fluviale del Nilo e una regione vulnerabile ai cambiamenti climatici e ad un elevato stress idrico dal momento che circa meta del bacino e un’area arida dove la popolazione dipende per il 93% dall’acqua dolce del fiume Nilo. Considerando che attualmente l’acqua del Nilo e enormemente utilizzata al di sopra della sua disponibilità e che il cambiamento climatico aumenta il rischio dell’insicurezza idrica e alimentare degli stati circostanti, si pone la necessita di individuare un indice per identificare l’occorrenza degli eventi di magra, che sia efficiente ed in grado di supportare i diversi decisiori ed operatori locali nella gestione delle risorse idriche. Questo studio si concentra sulla progettazione di un indice di magra specifico per il bacino del fiume Nilo, caso studio caratterizzato da un’eterogeneità climatica che varia da tropicale ad arido. Inoltre, il bacino del Nilo e frutto di discussioni internazionali per la critica gestione (finora non coordinata) delle risorse idriche tra i paesi limitrofi. La progettazione dell’indice viene effettuata attraverso l’applicazione del Framework for Index based Drought Analysis (FRIDA) sviluppato da Zaniolo et al. (2018). Il framework si basa su 3 fasi principali: recupero delle informazioni, estrazione delle caratteristiche e costruzione dell’indice. La prima consiste nella definizione della variabile target e nella raccolta dei predittori di input; la variabile target e stata scelta come Normalized Drought Vegetation Index (NDVI), poiché l’economia dei paesi del bacino e fortemente dipendente dall’agricoltura, ciò evidenzia la caratteristica della variabile target per rilevare le siccità agricole oltre a quelle meteorologiche, mentre le variabili di previsione sono state scelte in base all’elevata correlazione con gli eventi di siccità (i.e., precipitazioni e temperatura), oltre alla portata fluviale a causa del basso flusso del fiume Nilo e all’evapotraspirazione in quanto variabile critica causata dall’elevata evaporazione soprattutto nel Lago di Nasser, mentre l’umidita del suolo e stata considerata per tenere conto dell’effetto a lungo termine di una siccità. La seconda fase e stata eseguita utilizzando Wrapper for Quasi-Equally Informative Subset Selection (W-QEISS) come strumento per l’estrazione delle caratteristiche, in base alle prestazioni relativamente migliori dei wrapper rispetto ai filtri. La terza fase di costruzione del nuovo indice e stata eseguita attraverso due diversi modelli di regressione (lineare e a reti neurali artificiali), che permettono di confrontare le prestazioni dei due modelli scelti, in modo tale che i modelli siano caratterizzati uno per la semplicità e l’altro per l’elevata accuratezza predittiva quando vengono applicati a sistemi complessi. Dai risultati ottenuti si può osservare l’importanza della suddivisione dell’area di studio del bacino in base alla risoluzione spaziale dei dati di input raccolti, in modo che la procedura garantisca il filtraggio del rumore e allo stesso tempo eviti un sistema complesso di grandi dimensioni. I risultati dello studio hanno mostrato un’ottima performance degli indici di magra specifici per il bacino nel riprodurre la variabile target, dove l’accuratezza dell’indice nel 90% dei sottobacini e risultata molto alta (nella maggior parte dei casi con un coefficiente di determinazione superiore al 85%). D’altra parte, lo studio ha fornito una conferma dell’efficienza di W-QEISS nel raggiungere l’obiettivo del framework proposto, dove si e costruito un indice di magra sulla base delle caratteristiche predittive selezionate più accurato di quello che si sarebbe ottenuto utilizzando tutti i predittori di input.
A machine learning framework to design basin specific drought indexes
MIAARI, SAMI
2021/2022
Abstract
Droughts can have a serious impact on health, agriculture, economy, energy and environment. Monitoring and forecasting such phenomena is an essential task for mitigation, adaptation, and reduction of future risks by policymakers, mainly for areas that are characterized by high water demand with respect to the available fresh water. The Nile River Basin is a region vulnerable to climate change facing high water stress, since almost half of the basin is an arid area, where the population relies on 93$%$ of their fresh water from the Nile River. Considering that the Nile waters are fully exploited to fulfill the high water demands downstream, climate change will exacerbate the water-related challenges, leading to an increased risk of water scarcity and food insecurity. This raises the necessity of a reliable drought detection index to support water managers and decision makers in avoiding such risks. Thus, this thesis focuses on the design of a basin-specific drought index over the Nile River Basin, which consists in an interesting case study characterized by climate and topographical heterogeneity. Also, the Nile transboundary waters are cause of conflicts among the sharing countries and still generate international discussion. The index design is performed through the application of a machine learning FRamework for Index based Drought Analysis: FRIDA cite{zaniolo2018automatic}. The proposed framework was applied on the sub-basin scale, due to the complexity and heterogeneity of the system. Thus, the Nile River Basin is divided into 10 sub-basins according to hydrology, climate, and land cover. The framework is based on 3 main steps: information retrievement, feature extraction, and index construction. The first is performed by the definition of a target variable and collection of prediction variables: the Normalized Drought Vegetation Index is chosen as the target variable, since the economy of the basin countries is highly dependent on agriculture, highlighting the ability of the target variable in detecting agricultural and meteorological droughts. The prediction variables are chosen based on literature, identifying the high correlation with drought events, long term effect, and case study characteristics. The second step is performed by using Wrapper for Quasi-Equally Informative Subset Selection as the tool for feature extraction according to the relatively better performance of wrappers than filters. The third step---index construction---performed using two alternative regression models (i.e., linear and Artificial Neural Networks). A common thread highlighted from the output results is the consistency of the study area basin subdivision criteria, however an additional criteria of considering the spatial resolution of the input variables can be an improvement, since the sub-basins have different surface areas, topographic, and climatic characteristics. Thus, accounting for noise filtering, avoiding the effect of extremes and outliers in a portion of the sub-basin. Furthermore, the results convey high performance of the designed basin specific drought indexes in reproducing the target variable, where the index predictive accuracy in 90$%$ of the sub-basins is outstanding; while the low performance in some cases is due to multiple factors, mainly land cover and surface area. The study contributes a further proof of the expected feature extraction efficiency in meeting the objectives of the proposed framework, reproducing a drought index for the Nile River Basin highly accurate in most of the cases. The application of FRIDA can provide an appropriate drought monitoring index that can adapt to basins with different characteristics.| File | Dimensione | Formato | |
|---|---|---|---|
|
2022_07_Miaari.pdf
accessibile in internet per tutti
Descrizione: Thesis
Dimensione
7.24 MB
Formato
Adobe PDF
|
7.24 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/10589/189748