Flood Susceptibility Mapping Using a Deep Neural Network Model: The Case Study of Southern Italy ^†

Abstract

This study suggests a rapid methodology to delineate areas prone to flood using machine learning techniques. Based on available historically flooded areas, the model employs and combines globally collectible and reproducible conditioning factors to analyze flood susceptibility. The flood inventory map includes historically flooded areas from 1920 that occurred over the study area—Southern Italy. The impact of each factor is examined using correlation attribute evaluation and information gain ratio, while the performances of the model are evaluated by using area under receiving operating characteristics. Findings demonstrate that machine learning models can help in quick flood-prone areas analysis, especially in areas where flood hazard maps are not available.

1. Introduction

Flood hazard maps influence risk reduction and appropriate land use planning. In Italy, they are traditionally obtained through hydrologic–hydraulic modelling. However, due to the complexity of natural phenomena, even the most detailed models adopt simplifications to reproduce the inundation process [1]. The lack of available economic resources does not allow the study of all the critical cases on the territory. Therefore, flood hazard studies are usually limited to main rivers or specific tributaries, leaving a significant part of the territory unclassified. The spreading of machine learning applications, together with the acquisition of data from multiple sources, has given new potential to flood risk mapping, allowing the identification of areas most susceptible to flooding based on geomorphological characterization [2,3,4,5]. To achieve flood risk assessment at a large scale, this study suggests a flood susceptibility mapping based on machine learning techniques in a GIS (Geographic Information System) environment. The procedure is applied to the homogeneous physiographic units into which the study area is divided. For each unit, the model identifies flood-prone areas by analyzing the relationships between the past flood events and their environmental conditioning factors. Thus, flood modelling is formulated as a binary pattern recognition problem in which areas are divided into prone and no-prone to flood. The proposed approach allows a first identification of critical studies, which can then be the subject of more detailed analysis, leading to a better complementary understanding of the existing flood hazard maps, which remain the official normative reference.

2. Materials and Methods

2.1. Study Area

The study area (∼84,338 km²) includes five regions in Southern Italy: Basilicata, Calabria, Campania, Puglia, and Sicilia. It coincides with the area of interest of “PON Governance e Capacità Istituzionale 2014–2020” project, developed by the Italian Civil Protection National Department (http://governancerischio.protezionecivile.gov.it/pon-governance accessed on 1 July 2022). The study area was divided into homogenous territorial units (Figure 1). Starting from the existing scientific literature on territorial subdivision (altimetric [6], idro-morphologic [7], and geomorphologic [8]) the choice fell on the classification proposed by ISPRA (Istituto Superiore per la Protezione e la Ricerca Ambientale) in the so-called «Carta dei Tipi e delle Unità Fisiografiche dei Paesaggi Italiani» [9], part of the “Nature Map” 1:250,000 scale, as one of the most complete methods to identify homogenous areas. This use in the field of flood susceptibility is a promising innovation [10].

2.2. Flood Inventory Map and Flood Conditioning Factors

A significant effort was devoted to the construction of a database of the available records of flooding episodes over the study area to derive a map of historically flooded area. A total of over 200 flood events, corresponding to almost 1259 km² of historically inundated areas, have been collected (PON Governance e Capacità Istituzionale 2014–2020—Attività A21 “Potenziamento dei sistemi di previsione e di allertamento: mappatura della pericolosità dei fenomeni idrogeologici e idraulici non analizzati nei PAI e/o PGRA”). The historical flooded areas were converted into a binary map, assigned code 1 for flooded areas and code 0 for potentially non-floodable areas. The latter are defined as the external part of the so-called “maximum potentially floodable areas”, which derive from the union of flooded areas, flood hazard maps, alluvial deposits, slope (<30°) and 7 out of 10 most flooded types of geomorphons [8]. An extensive review of the literature was useful to extract the 13 most influential conditioning factors on flooding, whose ranking was based on an “Analytical Hierarchy Process” (AHP), a common multi-criteria decision-making technique [11]. These are: geomorphons, land cover (CLC), hydrologic soil group (HSG), curve number (CN), physiographic types of landscape, curvature, elevation, distance to the river, relative elevation, geomorphic flood index (GFI), slope, stream power index (SPI), and topographic wetness index (TWI).

2.3. Deep Neural Network

We employed a statistical model, namely TabNet (Attentive Interpretable Tabular Learning) [12]. It is an algorithm of a deep neural network (DNN) developed by the research department of Google Cloud AI (https://cloud.google.com/blog/products/ai-machine-learning/ml-model-tabnet-is-easy-to-use-on-cloud-ai-platform accessed on 1 July 2022), with an open-source version directly implementable in Python (https://github.com/dreamquark-ai/tabnet accessed on 1 July 2022). The model employs conditioning factors to predict flood susceptibility in a certain zone, providing an index between 0 and 1, which is interpreted as a probability of the zone to be prone to flood. To avoid overfitting, the dataset was partitioned into 3 sets: training set (used to train the model), validation set (used to adjust the hyperparameters) and testing set (used to evaluate generalization ability).

2.4. Performance Criteria

The flood susceptibility analysis implemented in the study defines a hierarchy based on the predisposing factors’ relative importance in flood triggering for each physiographic unit. This importance, also called “Average Merit”, was examined using correlation attribute evaluation and information gain ratio, analyzing the relation between observations and conditioning factors [13]. Additionally, a correlation analysis (both Pearson and Spearman) between the predisposing factors was made to avoid the repetition and the redundant increase of variables, which lead to a rise of the model’s complexity and its instability. The quality of each binary classifier was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curves that represent a good measure of performance, widely used in susceptibility analysis.

3. Results and Discussion

Table 1. Best predisposing factors’ combinations for each territorial unit.

Physiographic Unit	Flood Conditioning Factors	Non-Linear Correlations	Testing Accuracy	Testing AUC
Lowland	SPI, GFI, relative elevation, HSG	/	0.984	0.998
Tableland	Physio. types, relative elevation, CN, TWI	relative elevation—TWI	0.981	0.998
Hillside	GFI, relative elevation, elevation, HSG	/	0.966	0.995
Mountain	elevation, CLC, GFI, TWI	/	0.982	0.999
Islands	/	/	/	/

shows the results of the testing phase, the combination of the four most important predisposing factors for each territory. The conclusion of this analysis coincided with the elaboration of a susceptibility flood index that associates an “a priori” probability based on morphological characteristics of each single territorial element, as shown in Figure 2.

4. Conclusions

The results of the flood susceptibility mapping through the TabNet neural network are therefore consistent both with the inventoried inundations events and with the delimitation of hydraulic hazard maps. The method therefore constitutes an efficient way to identify the flood-prone areas, a useful tool for civil protection planning and a promising integration to evaluate the hydrogeologic risk. We suggest a deepening of this important and underestimated issue, in order to build classes concretely related to the flood susceptibility characteristics of the area. The ability to correctly classify the territory is found to be heavily related to the delineation of the potentially non-floodable areas (0), although their proper definition is often a neglected issue. Therefore, future developments may include alternative methodologies (such as marginal hazard areas) as part of the creation of the potentially non-floodable areas.