The increasing attention to the risks arising from floods requires a deeper and more comprehensive knowledge of the physical phenomena that govern them. In the framework of the program “Programma per il supporto al rafforzamento della governance in materia di riduzione del rischio ai fini di protezione civile” implemented by the Civil Protection Department within the NOP Governance and Institutional Capacity 2014-2020 and co-financed by the European Union, specific attention was paid to flood phenomena not studied within the current mapping of hydraulic hazard (mainly focused on the primary hydrographic network due to limited economic availability). In the project, a flood susceptibility analysis was carried out for five southern Italian regions (Basilicata, Calabria, Campania, Apulia, Sicily) belonging to the Southern Apennine District Basin Authority and to the Basin Authority of the Hydrographic District of Sicily, as an integrative tool to produce hydraulic risk scenarios for civil protection purposes. The susceptibility map derives from the union of four different deep neural network (DNN) models, called TabNet (Attentive Interpretable Tabular Learning), specific for structured data. Each model corresponds to a homogeneous territorial unit (HTU), identified in the physiographic areas of the “Carta dei Tipi e delle Unità Fisiografiche dei Paesaggi Italiani” at the scale 1:250000, whose use is a promising novelty. In fact, the physiographic areas have a clear degree of susceptibility: Lowland (maximum susceptibility), Tableland, Hillside, Mountain (minimum susceptibility). Susceptibility analysis has the advantage of using easily available conditioning factors, allowing a mapping of the whole area of the NOP (including the secondary network), here reported according to four classes of flood susceptibility (1 = low, 2 = medium-low, 3 = medium-high, 4 = high) and with a spatial resolution of 20 m. For the Lowland, where 68.75% of the historical floods fall, the map is also returned at 5 m resolution. The class 4 territory, which should have a priority role in the allocation of economic resources for more detailed hydraulic studies, also allows the identification of the most susceptible regions: Apulia, Campania, Sicily, Calabria, Basilicata. The latter sequence is also in agreement with the extension of the P1 and P2 hydraulic hazard areas (mosaicked by ISPRA in 2017) on a regional basis, except for the exchange between Sicily and Calabria (whose distinction is not clear-cut). In addition, the map is also defined for altitudes > 350 m, unlike previous mappings based on bayesian methods (Weight Of Evidence – WOE) by Menduni et al. (2020b). The comparison with the WOE method shows how TabNet can classify a higher percentage of historically flooded or floodable territory (P1, P2, P3) in class 4, leading to a better classification with increasing elevation (especially in physiographic areas with limited historic flood inventory, such as Mountain). This result is obtained by selecting only the four most important and uncorrelated factors for each HTU as input to the TabNet models, reducing both computational time and overfitting. The relative importance of factors is determined by both site-aspecific (Analytical Hierarchy Process – AHP) and site-specific (Correlation Attribute Evaluation – CAE and Information Gain Ratio – IGR) techniques. Only the latter can describe the spatial specificity of the four HTU, providing for each one different ranking’s results of importance. Linear correlation was assessed by Pearson's coefficient, while “non-linear” correlation was assessed by Spearman's rank correlation. In addition, susceptibility mapping, which often consists in a binary classification problem, is made providing the model with both historically flooded areas and/or areas of hydraulic hazard (“1”) along with areas that are neither flooded nor reasonably likely to be flooded in the future (“0”). A consistent spatial delimitation “0” pixel (which is one of the main reasons of such high AUC values) is made by inverse masking of the “maximum potentially flooded areas”. Nevertheless, the delineation of “0” areas is often neglected in the flood susceptibility literature, so it represents one of the future investigation branches in this field. Further future developments could be the comparison between TabNet and Convolutional Neural Networks (CNN) models, since CNN could be able to reduce the need for feature engineering (such as correlation-based methods, IGR or multi-collinearity analysis), and the definition of the correct resolution of the susceptibility map (i.e. 5 or 20 m) depending on the scale of representation of conditioning factors.
La crescente attenzione riguardo ai rischi da alluvione richiede una conoscenza più approfondita dei fenomeni fisici che li governano nella loro connessione con le caratteristiche geomorfologiche delle aree sui quali insistono. Nell’ambito del “Programma per il supporto al rafforzamento della governance in materia di riduzione del rischio ai fini di protezione civile”, realizzato dal Dipartimento della Protezione Civile nell’ambito del PON Governance e Capacità Istituzionale 2014-2020 e co-finanziato dall’Unione Europea, si è posta una specifica attenzione sui fenomeni alluvionali non studiati all’interno delle attuali perimetrazioni di pericolosità idraulica (specialmente sul reticolo idrografico primario a causa delle limitate disponibilità economiche). Pertanto, nel progetto si è portata avanti, per cinque regioni del Sud Italia (Basilicata, Calabria, Campania, Puglia, Sicilia) afferenti all’Autorità di Bacino Distrettuale dell’Appennino Meridionale e all’Autorità di Bacino del Distretto Idrografico della Sicilia, un’analisi di suscettibilità d’alluvione, come strumento integrativo per la produzione di scenari di rischio idraulico ai fini di protezione civile. La mappa di suscettibilità deriva dall’unione di quattro diversi modelli di rete neurale profonda (DNN), chiamata TabNet (Attentive Interpretable Tabular Learning), specifica per dati strutturati. Ogni modello corrisponde ad un’unità territoriale omogenea (UTO), individuata negli ambiti fisiografici della “Carta dei Tipi e delle Unità Fisiografiche dei Paesaggi Italiani” alla scala 1:250000, il cui utilizzo, con questo lavoro, si mostra assai promettente. La suscettibilità ha il vantaggio di usare fattori predisponenti di facile reperibilità, permettendo una mappatura dell’intera area del PON (compreso il reticolo secondario), qui riportata secondo quattro classi di suscettibilità alluvionale (1 = bassa, 2 = medio-bassa, 3 = medio-alta, 4 = alta) e con una risoluzione pari a 20 m. Per la Bassa Pianura, in cui ricade il 68.75% delle alluvioni storiche, la mappa è restituita anche a 5 m. Il territorio in classe 4, su cui potranno essere allocate prioritariamente le risorse per studi idraulici di maggior dettaglio, permette anche l’individuazione delle regioni più suscettibili: Puglia, Campania, Sicilia, Calabria, Basilicata. La sequenza trovata è anche concorde con l’estensione delle aree a pericolosità idraulica P1 e P2 (mosaicate da ISPRA nel 2017) su base regionale, tranne per lo scambio tra Sicilia e Calabria (la cui distinzione non è netta). Inoltre, la mappa è definita anche per quote > 350 m, a differenza delle precedenti mappature tramite metodi bayesiani (Weight Of Evidence – WOE) di Menduni et al. (2020b). Il confronto con il WOE mostra come TabNet riesca a classificare in classe 4 una maggiore percentuale di territorio storicamente alluvionato o alluvionabile (P1, P2, P3), portando ad una migliore classificazione all’aumentare della quota (soprattutto in zone con inventario alluvionale limitato, come in Montagna). L’elevate prestazioni del modello TabNet sono ulteriormente verificate dal superamento del benchmark di letteratura in termini di Area Under Curve (AUC) (98.65% in Hong et al. (2018b)), mostrando un valore minimo di AUC (del testing set) pari al 99.48% in Collina e massimo pari al 99.90% in Montagna. Questo risultato è ottenuto, selezionando in ingresso ai modelli TabNet, solo i 4 fattori più importanti per ogni UTO ed incorrelati tra di loro, riducendo sia i tempi di calcolo che l’overfitting. Il ranking d’importanza relativa dei fattori è estratto sia da tecniche sito-aspecifiche (Analytical Hierarchy Process – AHP), sia sito-specifiche (Correlation Attribute Evaluation – CAE e Information Gain Ratio – IGR). Solo quest’ultime riescono a descrivere la specificità territoriale delle quattro UTO, fornendo per ognuna dei risultati d’importanza relativa diversi. La correlazione lineare è stata valutata tramite il coefficiente di Pearson, mente quella “non lineare” tramite la correlazione per ranghi di Spearman. Inoltre, la suscettibilità, riconducendosi spesso ad un problema di classificazione binaria, è costruita fornendo al modello sia le aree storicamente alluvionate e/o aree a pericolosità idraulica (“1”), sia le aree non alluvionate né ragionevolmente alluvionabili in futuro (“0”). Una coerente definizione dei pixel “0” (anch’essa alla base degli elevati valori di AUC raggiunti) è operata tramite un mascheramento inverso delle “massime aree potenzialmente allagabili”. Nonostante ciò, la delimitazione delle aree “0” è spesso trascurata in letteratura, pertanto rappresenta uno dei futuri filoni d’indagine nell’ambito della suscettibilità d’alluvione. Ulteriori sviluppi futuri potrebbero essere il confronto tra TabNet e le reti neurali convoluzionali (Convolutional Neural Network – CNN), le quali possono ridurre la necessità d’estrazione delle caratteristiche dai dati (feature engineering, come i metodi basati sulla correlazione, IGR o analisi della multi-collinearità), e la definizione della corretta risoluzione di restituzione della mappa di suscettibilità (i.e. 5 o 20 m) in funzione della scala di rappresentazione dei fattori predisponenti.
Analisi di suscettibilità da alluvione tramite l'uso di Deep Neural Networks. Un caso studio su cinque regioni italiane
Balestra, Filippo
2020/2021
Abstract
The increasing attention to the risks arising from floods requires a deeper and more comprehensive knowledge of the physical phenomena that govern them. In the framework of the program “Programma per il supporto al rafforzamento della governance in materia di riduzione del rischio ai fini di protezione civile” implemented by the Civil Protection Department within the NOP Governance and Institutional Capacity 2014-2020 and co-financed by the European Union, specific attention was paid to flood phenomena not studied within the current mapping of hydraulic hazard (mainly focused on the primary hydrographic network due to limited economic availability). In the project, a flood susceptibility analysis was carried out for five southern Italian regions (Basilicata, Calabria, Campania, Apulia, Sicily) belonging to the Southern Apennine District Basin Authority and to the Basin Authority of the Hydrographic District of Sicily, as an integrative tool to produce hydraulic risk scenarios for civil protection purposes. The susceptibility map derives from the union of four different deep neural network (DNN) models, called TabNet (Attentive Interpretable Tabular Learning), specific for structured data. Each model corresponds to a homogeneous territorial unit (HTU), identified in the physiographic areas of the “Carta dei Tipi e delle Unità Fisiografiche dei Paesaggi Italiani” at the scale 1:250000, whose use is a promising novelty. In fact, the physiographic areas have a clear degree of susceptibility: Lowland (maximum susceptibility), Tableland, Hillside, Mountain (minimum susceptibility). Susceptibility analysis has the advantage of using easily available conditioning factors, allowing a mapping of the whole area of the NOP (including the secondary network), here reported according to four classes of flood susceptibility (1 = low, 2 = medium-low, 3 = medium-high, 4 = high) and with a spatial resolution of 20 m. For the Lowland, where 68.75% of the historical floods fall, the map is also returned at 5 m resolution. The class 4 territory, which should have a priority role in the allocation of economic resources for more detailed hydraulic studies, also allows the identification of the most susceptible regions: Apulia, Campania, Sicily, Calabria, Basilicata. The latter sequence is also in agreement with the extension of the P1 and P2 hydraulic hazard areas (mosaicked by ISPRA in 2017) on a regional basis, except for the exchange between Sicily and Calabria (whose distinction is not clear-cut). In addition, the map is also defined for altitudes > 350 m, unlike previous mappings based on bayesian methods (Weight Of Evidence – WOE) by Menduni et al. (2020b). The comparison with the WOE method shows how TabNet can classify a higher percentage of historically flooded or floodable territory (P1, P2, P3) in class 4, leading to a better classification with increasing elevation (especially in physiographic areas with limited historic flood inventory, such as Mountain). This result is obtained by selecting only the four most important and uncorrelated factors for each HTU as input to the TabNet models, reducing both computational time and overfitting. The relative importance of factors is determined by both site-aspecific (Analytical Hierarchy Process – AHP) and site-specific (Correlation Attribute Evaluation – CAE and Information Gain Ratio – IGR) techniques. Only the latter can describe the spatial specificity of the four HTU, providing for each one different ranking’s results of importance. Linear correlation was assessed by Pearson's coefficient, while “non-linear” correlation was assessed by Spearman's rank correlation. In addition, susceptibility mapping, which often consists in a binary classification problem, is made providing the model with both historically flooded areas and/or areas of hydraulic hazard (“1”) along with areas that are neither flooded nor reasonably likely to be flooded in the future (“0”). A consistent spatial delimitation “0” pixel (which is one of the main reasons of such high AUC values) is made by inverse masking of the “maximum potentially flooded areas”. Nevertheless, the delineation of “0” areas is often neglected in the flood susceptibility literature, so it represents one of the future investigation branches in this field. Further future developments could be the comparison between TabNet and Convolutional Neural Networks (CNN) models, since CNN could be able to reduce the need for feature engineering (such as correlation-based methods, IGR or multi-collinearity analysis), and the definition of the correct resolution of the susceptibility map (i.e. 5 or 20 m) depending on the scale of representation of conditioning factors.File | Dimensione | Formato | |
---|---|---|---|
TESI_Filippo-Balestra.pdf
Open Access dal 12/04/2023
Descrizione: TESI_Filippo-Balestra
Dimensione
50.2 MB
Formato
Adobe PDF
|
50.2 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/10589/187572