Riverine flood potential assessment using metaheuristic hybrid machine learning algorithms

This study presents the performance of stand‐alone and novel hybrid models combining the feed‐forward neural network (FFNN) and extreme gradient boosting (XGB) with the genetic algorithm (GA) optimization to determine the riverine flood potential at a local spatial scale, which is represented by the Gidra river basin, Slovakia. Eleven flood factors and a robust flood inventory database, consisting of 10,000 flood and non‐flood locations, were used. Using the FFNN, XGB, GA‐FFNN and GA‐XGB models, 16.5%, 11.0%, 17.1%, and 12.3% of the studied basin, respectively, is characterized with high to very high riverine flood potential. The applied models resulted in very high accuracy, that is, AUC = 0.93 in case of the FFNN stand‐alone model and AUC = 0.96 in case of the XGB stand‐alone model. The GA algorithm was able to raise the value of AUC for the hybrid GA‐FFNN and GA‐XGB models to 0.94 and 0.97, respectively. The results of this study can be useful, especially, for the identification of the areas with the highest potential for riverine floods within the next updating of the Preliminary Flood Risk Assessment, which is being carried out based on the EU Floods Directive.


| INTRODUCTION
Riverine flooding is one of the most frequent type of flooding, which is caused by long-term rainfall and gradual increase of water level, which affects the river bed and adjacent floodplain. The mapping and assessment of riverine flood potential focuses on defining the propensity of an area for fluvial flooding based on the assessment of various flood factors, which are usually represented by physical-geographical factors, such as slope, river density, lithology, soils, and land use/land cover (LULC) (Costache, 2019a;Costache et al., 2021;Jacinto et al., 2015;Vojtek et al., 2022Vojtek et al., , 2023Zaharia et al., 2015Zaharia et al., , 2017.
There are various physical-geographical factors, which can be used to determine the level of flood potential in a particular area, and these can be generally divided into morphometric factors, hydrographic factors and factors influencing the infiltration of precipitation. Morphometric factors are represented, for example, by elevation, slope angle, curvature or various indices, like topographic wetness index (TWI), topographic position index (TPI), stream power index (SPI), sediment transport index (STI), which are actually secondary derivates of slope or flow accumulation (Beven & Carling, 1989;Malekinezhad et al., 2021). Hydrographic factors can be distance from river, flow accumulation or river density (Smith & Ward, 1998). Factors influencing the infiltration of precipitation and, thus runoff, usually include lithology, soil texture and land cover (Smith & Ward, 1998;Vojtek & Vojtekov a, 2018, 2019. In addition, climatic factors are also used, but in this study, we adhere to concept of flood potential, which does not count with rainfall or discharge data, that is, triggering factors, but focuses only on preliminary flood conditioning factors caused by morphometric, hydrographic and infiltration factors. One of the important tasks in flood potential mapping is to choose appropriate factors. This is usually done subjectively, as the selection depends on the type of flood being investigated, data availability and consistency and study area properties. In addition, it is recommended investigating the multicollinearity among independent factors in order to exclude the factors having high correlation between each other. In particular, multicollinearity methods may include Pearson correlation (Islam et al., 2021), variance inflation factor (Rahmati et al., 2020), and the like.
Besides the aforementioned independent flood conditioning factors, the key role is played by flood inventory, which consists of historical flood locations that are used for training and testing the chosen models. In this step, flood locations can be represented by single flood points (Bui et al., 2016;Chen et al., 2019;Wang et al., 2020) or seed of flood points , which can be derived from flood extent polygons. From these two options, the usage of seed of flood points seems to be more reasonable and reliable, as the representation of flood by a single point can be questionable (Al-Abadi & Pradhan, 2020).
In particular, few flood potential studies have attempted to optimize the ML models with the genetic algorithm (GA). For example, Linh et al. (2022) optimized the extreme gradient boosting (XGB) with the GA (area under curve-AUC = 0.87), Dodangeh et al. (2020) applied GA optimization with the group method of data handling (GMDH) (AUC = 0.72) and support vector regression (SVR) (AUC = 0.75) models, Arabameri et al. (2022) used hybridization of GA with the support vector machine (SVM) (AUC = 0.89) and particle swarm optimization (PSO) (AUC = 0.90) or Hong et al. (2018) used adaptive neuro fuzzy inference system (ANFIS) with GA (AUC = 0.88). Although the mentioned AUC in these studies is lower or equal 0.9, the capability of GA to improve the prediction is high. It is evidenced by Shahabi et al. (2021), who proposed the hybrid model of deep belief network with back propagation (BP) algorithm and the GA, which resulted in the AUC = 0.99, or Arora et al. (2021), who applied ANFIS with GA resulting in AUC = 0.92.
This study aims to present hybrid models combining FFNN and XGB with GA optimization to determine the level of riverine flood potential at a local spatial scale, which is represented by the Gidra river basin, Slovakia. The main novelty of this study is seen in the hybridization of FFNN with GA, which has not been used so far for flood potential prediction, and its comparison with hybrid GA-XGB model and FFNN and XGB stand-alone models. The local spatial scale used in this study assures the high precision and general consistency of source data, which is very important for the reliability and practical applicability of the model results.

| STUDY AREA
The study area is represented by the Gidra River Basin (western Slovakia), which has a total area of 198 km 2 . The main river is the Gidra River, which has a total length of 38.6 km and it mouths into the Dolný Dudv ah River as its right-sided tributary. The main tributaries of the Gidra River are theŠtefanovský potok (stream length of 11.4 km) and Ronava (stream length of 17 km). Geo According to the geomorphological division of Slovakia, the upper part of the basin is formed with Malé Karpaty (mountain) while the majority of the basin belongs to Podunajsk a pahorkatina (hills) and a small area near the outlet is included in the Podunajsk a rovina (plain) (Mazúr and Lukniš, 1986). The elevation difference is 575 m and the average slope is 6.2%.
The Gidra River Basin lies in the temperate climatic zone. The average annual temperature in the studied basin is 9.5 C. The mean annual precipitation is approximately 700 mm for the upper part of the basin while for the lower part it is 564 mm (Bochníček et al., 2015).
LULC of the basin is dominated by arable land (63.5%). The share of forests, glades, shrub, and urban greenery on the basin area is 24.8%. In case of built-up areas, backyard gardens, and dumps the share is 6.5%, grasslands (2.7%), vineyards and orchards (1.2%), watercourses and water bodies (0.7%), and railway and roads (0.6%).
According to the Nomenclature of Territorial Units for Statistics (NUTS) classification (https://ec.europa.eu/ eurostat/web/nuts/background), the study area is included in NUTS II Western Slovakia, NUTS III Trnava Region and NUTS III Bratislava Region.
Several historical flood events occurred in the basin, for example, riverine floods in July 1997, April 2006, March 2009, May 2010, September 2014, July 2016 or heavy flash flood on 7th June 2011. Altogether, 10 municipalities and their urbanized area, through which the Gidra River flows, were affected by historical floods: Píla,Čast a,Štefanov a, Budmerice, Cífer, Jablonec, Slovensk a Nov a Ves, Voderady, Mal a Mača, and Pavlice.

| Flood conditioning factors
The data on flood conditioning factors represent basic input for potential mapping. It has to be noted that there are no general principles for the selection of flood conditioning factors and this task is usually carried out subjectively. Furthermore, selection of flood factors depends on the availability of data as well as on the type of flood to be assessed, e.g. pluvial or riverine flood. Altogether, 11 flood conditioning factors were considered being relevant for the riverine flood potential analysis in the study area. Sources for the flood conditioning factors used are provided in Table 1.
Using the DEM (DMR3.5) with 10 m resolution, five morphometric factors of elevation, slope angle, plan curvature, profile curvature, TWI and SPI were derived ( Figure 2). Elevation was derived from the DEM using the ArcGIS software. Slope was calculated using the Slope tool in ArcGIS software and it identifies the steepness at each cell of a raster surface (DEM), that is, the lower the slope values, the flatter the terrain or the higher the slope value, the steeper the terrain. The plan (planform) and profile curvature were computed via the Curvature tool in ArcGIS software. The plan curvature is perpendicular to the direction of the maximum slope. It relates to the convergence and divergence of flow across a surface. A positive value indicates that the surface is laterally convex at that cell while a negative value indicates that the surface is laterally concave at that cell. Moreover, the value of zero means that the surface is linear. On the other hand, profile curvature is parallel to the slope and indicates the direction of maximum slope. It affects acceleration and deceleration of flow across the surface. A negative value indicates that the surface is upwardly convex at that cell and flow will be decelerated. A positive values indicates that the surface is upwardly concave at that cell and the flow will be accelerated. Similarly as in plan curvature, a value of zero indicates that the surface is linear. Factors of TWI and SPI were calculated in Arc-GIS software using Equation (1) and Equation (2) (Chowdhuri et al., 2020;Hreško et al., 2003).
F I G U R E 1 Study area-Gidra river basin.
where A s is the specific contributing area, that is a parameter of the tendency to receive water, and β is the slope gradient, that is a parameter of the tendency to evacuate the water (see Mattivi et al., 2019). Hydrographic factors include the drainage density, that is, the ratio of total length of all watercourses in a basin to the basin area calculated with the use of Line Density tool in ArcGIS software, and distance from rivers, that is, the distance from the line representing the watercourse calculated with the use of Euclidean Distance tool in ArcGIS software. The input data for both rasters was the vector layer of river network ( Figure 3).
Factors influencing the infiltration of water are represented by lithology, soils and LULC Lithological factor was derived based on the vector Map of Engineering Geological Zones Map at a scale of 1:50,000. Lithological units are presented in Figure 4. Soil factor was derived using the two vector maps: (i) bonited soilecological units (the so-called BPEJ), which include soil texture types on agricultural land, (ii) forest soil unit, which include soil texture types on forest land. Land use/land cover (LULC) factor was obtained from the Basic Data Base for the Geographic Information System (ZBGIS), which was supplemented by the visual interpretation of orthophotos from 2017 with the resolution of 25 cm.

| Flood inventory
In this study, the past floods used for training as well as testing the selected models, which are described in Section 3.2.2, were represented by flood extent polygon. This polygon was converted into a seed of points consisting of 10,000 flood points. The official documents of preliminary flood risk assessment and flood risk maps (Ministry of Environment of the Slovak Republic 2011, 2018) were used as sources for creating the flood inventory map. As for the analysis, flood points were randomly divided into 70% of training locations and 30% of testing locations ( Figure 1). Furthermore, the same number of 10,000 non-flood points was randomly created in ArcGIS software. Since the flood potential analysis focuses on riverine floods, the distance of non-flood points from river is higher than 400 m, which represents areas where the occurrence of riverine floods is the least probable in case of the presented river basin.

| Multicollinearity
The variance inflation factor (VIF) evaluates the intensity of multicollinearity among independent variables. The T A B L E 1 Source data used for deriving flood conditioning factors. shows how much of the change in the estimated coefficients for alignment has increased. Multicollinearity intensities can be analyzed by examining the magnitude of the VIF value. This index tests how the behavior of an independent variable changes due to its correlation with other independent variables. Therefore, by identifying the strength of the correlation, it facilitates the compatibility of the model, in the presence of high variance inflation, even small changes in the data or in the structure of the model (Alin, 2010;Tamura et al., 2019). The VIF index can be calculated from Equation (3): where R 2 represents the coefficient of determination of the variable. The VIF above 10 indicates a critical alignment position and a value close to 1 indicates a favorable position. The acceptable limit of VIF index is below 5. If the VIF test statistic is close to one, it indicates no alignment (Tamura et al., 2019).

| Machine learning models
Feed forward neural network The feed forward neural network (FFNN) model consists of a set of neurons in the input layer, usually selected by the number of input variables, and one or more layers in the hidden layers and one output layer. In this algorithm, the neurons of each layer are connected to the neurons of the previous layer, but this connection is not necessarily under the same conditions, but with different weights. The input to this network passes from one layer to another to reach the output layer and because there is no post-emission in it, it is known as the feed forward network (Bebis & Georgiopoulos, 1994). The architecture of these neural networks is such that all nodes are completely connected to each other. In these networks, activation from the input layer to the output is based on a hidden layer between the input and output without the F I G U R E 3 Hydrographic factors: (a) drainage density, (b) distance from river.
problem of backward loop. In most cases, the backpropagation approach is used to train the feed neural networks. The most important issue in the design of artificial neural networks is to determine the number of hidden layers and the number of neurons in them. The number of neurons directly affects the ability to train and induce neural networks. So that if the number is less, although the training time is shorter, but the network is not trained enough and it is not possible to get acquainted with all the modes. But, if this large number is selected, long computation time and larger storage capacity are required, but instead the training ability will be strengthened (Arulampalam & Bouzerdoum, 2003;Setiono, 2001). Schematic diagram of FFNN model is presented in Figure 5.

Extreme gradient boosting model (XGB)
The Extreme gradient boosting algorithm was introduced by Chen et al. (2015). This algorithm is constructed based on supervised ML, decision trees, ensemble learning, and gradient boosting concepts. XGB is an implementation of the Gradient Boosting concept, but what makes this algorithm unique is that it applies a more regular model formalization for over-fitting control, which gives it better accuracy. XGB can use any loss function that specifies a gradient. Due to parallel computing in this algorithm, it is one of the quick computing ML algorithms. The XGB method is used for regression and classification issues. This algorithm uses the booster algorithm to sequentially modify weak classification models and build them as strong classifiers. XGB operates by splitting data into segments that lead to accurate modeling depending on different parameters. The trees in the XGB take into account the previous prediction value for a given data point and make a new tree that divides the existing data as great as possible to maximize the 'gain' in prediction (Chen et al., 2015;Janizadeh et al., 2021). Schematic diagram of XGB model is presented in Figure 6.

| Optimization
GA are a family of computational models inspired by the concept of evolution. These algorithms are referred to as F I G U R E 5 Schematic diagram of FFNN model used.
F I G U R E 6 Schematic diagram of XGB model used.
function optimizers, that is, algorithm is used to optimize the objective functions of various problems. The implementation of a GA usually begins with the production of a population of chromosomes (the initial population of chromosomes in GAs is usually randomly generated and is bounded up and down by the problem variables) (Mirjalili, 2019). GA includes the operators of coding, evaluation, selection, composition, mutation, and decoding. First, the problem variables are encoded. In this step, the algorithm deals with the coded form of the parameters or variables instead of working on them. In the evaluation stage, the fitness function evaluates each string with a numerical value from the target function and determines its quality.
The higher the quality of the answer string, the greater the suitability of the answer, and the greater the likelihood of participating in the production of the next generation. In the selection step, a pair of chromosomes are selected to be combined. In fact, it is the operator of choosing the interface between two generations. The selection process is random and the criterion for selection is fitness. In a combination operator, the older generation of chromosomes mixes with each other to form a new generation of chromosomes. The combination operator in the GA eliminates the dispersion or genetic diversity of the population and actually allows good genes to be combined (Kumar et al., 2010;Li et al., 2018). The mutation operator generates other possible answers. In the mutation operator, with a certain probability, the strings of the previous generation are applied and the strings of the next generation of strings are added. The result of the operation of the mutation operator will increase the probability of escaping from local optimal points.
The decoding is the opposite of the encoding operator and after the algorithm has provided the best answer to the problem, the decoding is applied to the answers so that the answer appears as a real function. After generating each new set of binary strings, the fit value of each string is calculated and the necessary condition for termination of the algorithm is checked. If the condition considered for the implementation of the algorithm is not met, by repeating the above steps, new sets of strings are generated and the corresponding fit values are calculated. After achieving convergence or satisfying the termination condition, the algorithm is stopped and the best string obtained in the last generation is introduced as the optimal answer (Kumar et al., 2010;Li et al., 2018). Schematic diagram of GA optimization method is presented in Figure 7.

| Validation methods
The receiver operating characteristic (ROC) curve characterizes the receiver performance as a measure of efficiency in classification problems. The ROC curve is a graphical representation of the degree of sensitivity or true prediction versus false prediction (1-specificity) in a binary classification system in which the separation threshold varies. The area under the curve (AUC) shows the overall quality of the modeling (accuracy). The AUC is a number that measures one aspect of performance. The AUC value is between zero and one. The value of 0.5 displays that the model has random prediction and the value of 1 display that the model has an excellent prediction (Bradley, 1997). In this study, five criteria of ROC include Sensitivity, Specificity, Positive predictive value (PPV), Negative predictive value (NPV) and AUC for evaluating ML algorithms in modeling the riverine flood potential.
The five-criteria evaluation index were calculated from Equation (4): where TP (true positive) and TN (true negative) are accurately classified pixel numbers while FP (false positive) and FN (false negative) are falsely classified pixel numbers. P is the total number of riverine flood locations and N is the total number of riverine non-flood locations (Band et al., 2020;Mosavi et al., 2020).

| Results of multicollinearity
The variance inflation factor (VIF) was used for determining multicollinearity among 11 independent variables. The results of the VIF analysis are shown in Table 2. Based on this result, all 11 variables had VIF less than 5. The maximum inflation variance was observed with 3.43 in the elevation variable and the minimum with 1.05 in the SPI variable. Therefore, all these variables were used as the independent variables for modeling the riverine flood potential.

| Optimization of ML parameters
In order to optimize the FFNN and XGB ML algorithms parameters, the GA metaheuristic method was used. The result of optimization in the FFNN and XGB models is shown in Figure 8 and Figure 9, respectively. The results of optimal parameters in FFNN and XGB models are shown in Table 3 and Table 4, respectively. According to the result of optimization parameters in FFNN, the optimal size and decay parameters, based on GA method, were 8 and 0.67, respectively. Also, the results of determining optimal parameters in XGB show that optimal nround, lamba, alpha, and eta, based on GA optimization algorithm, were 511.80, À1.38, 0.53 and À4.82, respectively.

| Validation
In order to evaluate the accuracy of models for riverine flood potential modeling in this study, five different criteria in two-stage training and validation were used. The results of evaluation accuracy of four algorithms are shown in Table 5. In the training stage, all four algorithms had high accuracy in riverine flood potential modeling. Evaluation based on AUC showed that AUC in FFNN, GA-FFNN, XGB, and GA-XGB algorithms were 0.93, 0.94, 0.99 and 0.99, respectively. The results of evaluation accuracy of four ML algorithms based on AUC for the validation stage are shown in Figure 10. The GA optimization method increased the performance in two FFNN and XGB models. Moreover, all four ML algorithms have high performance in riverine flood modeling in the study area and the highest AUC-accuracy was recorded for GA-XGB model (AUC = 0.

| Riverine flood potential maps
The final maps of riverine flood potential, which were computed using the four applied models, are presented in Figure 11. The values of flood potential are shown on a continuous scale from 0 to 1. As can be seen in Figure 11, the areas with the highest potential to riverine flooding are naturally along the main channels of the Gidra River as well as theŠtefanovský potok.
Moreover, the extent of high and very high riverine flood potential, which corresponds to the values of 0.6-1.0 (see Figure 11), was compared to the flood extent in the official flood hazard map for Q 100 . This map was created by the Slovak Water Management Enterprise in 2015 under the EU Floods Directive requirements using the MIKE Flood model, 1D steady state flow conditions, and photogrammetrically created DEM. The official flood hazard maps can be found on the map portal of the Slovak Water Management Enterprise (SWME, 2022). Table 6 presents the results of comparison between the high and very high riverine flood potential (values of 0.6-1.0) with the flood extent in the official flood hazard map for Q 100 , which has the value of 9.47 km 2 .
As can be seen in Table 6, the best match with the flood extent in the official flood hazard map (Q 100 ) was recorded by the GA-XGB model (94.19%) and XGB model (94.09%), which are followed by the GA-FFNN model (89.12%) and FFNN model (87.54%). Results of the comparison between the modeled riverine flood potential and hydraulically modeled flood extent indicate a good correspondence and, therefore, practical applicability of the models used in this study, especially, the GA-XGB and XGB.

| Variable importance analysis
The results of variable importance analysis for riverine flood potential modeling based on four ML algorithms are shown in Table 7. Investigation of the variable importance analysis showed that independent variables have various effect on flood potential modeling in four ML algorithms. In the FFNN model, soil, drainage density and LULC variables had highest importance while in the GA-FNN model, elevation, drainage density and lithology variables had the highest importance. In case of the XGB and GA-XGB models, elevation, drainage density and distance from river variables had significant role in riverine flood potential modeling in the Gidra basin. Overall, the elevation variable had the highest importance in flood potential modeling based on GA-FNN, XGB, and GA-XGB models result while in the FFNN model, this variable had the lowest importance.

| DISCUSSION
In this study, two different types of ML models including FFNN and XGB models were applied for riverine flood potential modeling in the Gidra basin. The results showed that two ML models had high performance in modeling the riverine flood potential. However, evaluation of models in the training and validation stage showed that the XGB model had higher accuracy than the FFNN model. The XGB model integrates the prediction results of all basic learners, thus increasing its detection rate and generalizability. In addition, this model supports the custom loss function and adds a regular expression to the target function, which helps prevent over-fitting and simplifies the learning model, thereby increasing the learning rate (Chen et al., 2015). Ma et al. (2021) found out that the XGB-based method is an effective method for obtaining high-quality flash flood risk results also on the national level. Using the XGB model for modeling natural hazard phenomena in recent years proved that this model is one of the best ML methods to predict potential for floods, landslides and erosion (Chen et al., 2021;Janizadeh et al., 2021;Sahin, 2020). Furthermore, the GA metaheuristic algorithm was used for optimizing the parameters in the FFNN and XGB model. The results of optimization and building hybrid models for modeling riverine flood potential in  the Gidra basin showed that a combination of the FFNN and XGB model with GA decreases model error in the training and validation stage, that is, the accuracy of models increase. The most important strength of the GA is its parallelism so that there are several starting points for solving the problem and at one time, it can search the problem space from several different directions. This increases the efficiency of the GA in solving nonlinear problems with large space (Mirjalili, 2019). Due to the fact that flood is a non-linear and complex phenomenon, a change in one part may have an uncoordinated effect on the whole system or a change in several elements may have a great effect on the system. Therefore, parallelism of the GA solves this problem. In addition, the GA used for solving the problem shows random changes in the candidate solutions and uses the fit function to measure whether they have made progressive changes or not. This allows the algorithm to start solving the problem in a wider space and since its decisions are essentially random, all possible solutions are open to the problem (Kumar et al., 2010;Mirjalili, 2019). In recent years, various researchers used GA for modeling flood potential as a hyperparameter for optimization in different ML algorithms, such as ANFIS Hong et al., 2018), SVR (Dodangeh et al., 2020); Random Forest (Saleh et al., 2022), extreme gradient boosting (Linh et al., 2022;Saleh et al., 2022). The results showed that the usage of GA for optimizing ML algorithms increases the performance of the models. Limitations in the data and methods used pertain to the following: • It will be optimal if the data sources used for processing flood conditioning factors will have the same original map scale or spatial resolution.  Darabi et al. (2020). • When using the same methodology for other study area, the distance of non-flood points should be subjectively determined based on the studied river length, width, and depth and other characteristics of the river and study area. As long as our studied river (channel) is relatively small in terms of length, width, and depth, then there is the least probability that the riverine flood will occur in 400 m distance and more from the river, even in case of 1000-year flood scenario. Another limitation in the data is seen in the absence of actual inundated area derived from the field surveys or satellite images, which could have been used for validation purposes. • Regarding the transferability of the results to other regions, the accuracy of flood conditioning factors should be adjusted to the size of the study area, especially, the already mentioned DEM in previous point, but also other factors (like lithology, soils or LULC) in order that they have the best match in terms of spatial resolution or original map scale. Furthermore, the selection of non-flood points should be also adjusted based on the channel and floodplain parameters, especially, channel length, width and depth. The distance value of 400 m from the river used in this study has to be modified for the studied river in other regions. The applied methods, that is, four models, can be easily transferred also to other regions based on the adjusted input data.

| CONCLUSION
In this study, two stand-alone and two hybrid ML approaches were used for identification of areas with the potential to riverine flooding. Altogether, 11 flood conditioning factors and robust flood inventory database consisting of 10,000 flood and non-flood locations were used. Verification of the flood potential models confirmed very high accuracy with the values of AUC = 0.93 and 0.96 in case of the FFNN and XGB stand-alone models, respectively. The GA algorithm was able to raise the value of AUC for the hybrid GA-FFNN and GA-XGB models to 0.94 and 0.97, respectively. Recommendations from the performed research can be summarized as follows: • For flood potential studies, the general recommendation is to focus on mapping only one type of flood, which then implies the choice of relevant flood conditioning factors, for example, for potential riverine (fluvial) flooding, flash-flood flooding, pluvial flooding, etc.
• This study uses readily and publicly available data, which enable systematical mapping of riverine flood potential based on the spatial and temporal changes in chosen factors. • From the aspect of the applicability and reliability of the flood potential maps, it is recommended focusing on local spatial scale, where the most accurate modeling results can be achieved. • This study uses relatively consistent source data for riverine flood potential mapping, that is, differences in generalization of input data, such as the original map scale or spatial resolution, are very low. This is very important for reliable flood potential mapping with practical applicability. • Regarding the creation of flood inventory, it is recommended using the seed of flood points rather than single flood points, which may not represent the actual flood extent adequately.
The results of this study can be useful, especially, for the next updating of the Preliminary Flood Risk Assessment and the associated identification of geographical areas with potentially significant flood risk, which is being carried out in six-year cycles based on the EU Floods Directive.

FUNDING INFORMATION
This work was supported by the Scientific Grant Agency of the Ministry of Education, Science, Research and Sport of the Slovak Republic and the Slovak Academy of Sciences (VEGA) under the grant number 1/0103/22 through the project entitled "Spatio-temporal Changes and Prediction of Flood Risk in Municipalities of Slovakia".