On a global scale, floods constitute the major natural hazards causing the greatest incidence of damaged property as measured by the number of material damages caused by this phenomenon (Jacinto et al., 2015). The hazards associated with floods therefore pose a great challenge for each and every one of the authorities concerned with the flood risk management activity (Luino et al., 2018). As a result of the climatic changes that have taken place over the past few decades, it has become evident that floods have become more frequent and severe, but it has also contributed to a marked increase in the amount of material damage and human losses (Peng-cheng et al., 2016). This is one of the reasons why the number of scientific publications that address the topic of flood risk management has increased steadily in the past few years. Romania ranks second behind France in terms of the economic damage caused by floods at an EU level. Romania must, like all countries in Europe, ensure that its flood risk management program is in compliance with the Directive 2007/60/EC. This document suggests that, in order to manage flood risks in communities, a great deal of attention should be paid to identifying areas prone to flooding (Afriyanie et al., 2020). Furthermore, the evaluation of flood vulnerability is likewise one of the most critical non-structural measures implemented in order to reduce material losses and lives that result from natural disasters such as flooding (Arora et al., 2021). It is possible to quickly determine whether a given area is vulnerable to flooding by using the latest Geographic Information System technologies. It is clear that the efficiency of any GIS process will be determined, in a considerable measure, by the accuracy of its input data, as well as by the combination of GIS models with statistical and machine learning algorithms. A growing number of studies that have used GIS techniques in combination with some models specific to machine learning and bivariate statistics to identify and map flood-exposed zones, have been published in recent years (Ahmadlou et al., 2021). There are three bivariate statistical models most commonly used to develop models for evaluating flood susceptibility: Weights of Evidence, Statistical Index and Frequency Ratio (Sahana and Patel, 2019). There is a fundamental limitation with the bivariate statistical models because they consider only the spatial relation between the flood locations and the conditioning factors, but they do not consider the causal relationship between the predictors governing the flow of the floods (Costache, 2019a). Among the machine learning methods used to detect those surfaces showing flood prone characteristics, the most commonly used ones are: decision trees algorithms, adaptive neuro-fuzzy inference systems, support vector machines and artificial neural networks (Alizadeh et al., 2018; Khosravi et al., 2018; Xiong et al., 2019). One of the key advantages characteristics for machine learning models resides in the high automation that they provide, as well, as the ease with which they can identify patterns and trends into a data sample. Additionally, machine learning models can handle multiple diversity, as well as multi-dimensional, data sets (Elmahdy et al., 2020). Even though the machine learning models constitute the most advanced aspects of the data processing field, they also have certain disadvantages that make them less attractive. In the literature, there have been a great deal of discussions regarding two of the most significant drawbacks of this kind of technique: the large amount of data required and the high error susceptibility (Albon, 2018; Ao et al., 2019; Baghban et al., 2018).
In most cases, the authors of studies regarding assessment of flood potential for a particular region consider a multitude of geographic factors that influence the nature and severity of this natural disaster. Usually, the authors take into account the following flood predictors: rainfall, distance from rivers, soils, slope angle, land use, plan curvature, topographic wetness index and lithology (Chowdhuri et al., 2020). It is possible to compute the flood susceptibility by overlaying in a simple manner the flood in GIS software and analyzing their contributions. Advance models that allow the weighting of these factors are also able to determine flood susceptibility (Bui et al., 2020). As a general rule, flood conditioning factors are weighted according to their spatial relationships with historical flood occurrences (Azareh et al., 2019). As such, the locations of recent flood events are considered when giving local weights to these factors (Dano et al., 2019).
Taking into account these details, the main purpose of the present study is to identify the flood-prone areas in Prahova river's hilly and mountain basins, by using four models. Thus, the Deep Learning Neural Network (DLNN), Support Vector Machine (SVM) and also these 2 optimized models with Particle Swarm Optimization (PSO) were used to estimate the flood susceptibility.
As a part of the training of the models mentioned above, a number of 10 flood predictors, as well as 158 flood historical events, will be taken into account. One of the huge advantages of this approach is the possibility of evaluating and validating, quantitatively, the precision of the algorithms involved in the methodological workflow and also to assess the accuracy of the flood susceptibility maps. In the present study, the accuracy of the results has been tested by computing several statistical parameters and by plotting the ROC curve and calculating the area under the curve.