Search

Abstract In this article, we aim at the prediction of possible locations of already defunct historical objects with a defensive function (HODFs) in Slovakia, which have not been found and documented so far, using three machine learning methods. Specifically, we used the support vector machine, k-nearest neighbors, and random forest algorithms, which were trained based on the following five factors influencing the possible occurrence of HODFs: elevation, distance from a river, distance from a settlement, lithological rock type, and type of representative geoecosystems. Training and testing datasets were based on a database of already documented 605 HODFs, which were divided into 70% of training samples and 30% of testing samples. All of the three models reached the AUC-ROC value over 0.74 based on the testing dataset. The best performance was recorded by the random forest predictive model with the AUC-ROC value equal to 0.79. The results of the random forest model were also validated with the recently documented HODFs via the archeological research.

Scouring around bridge piers is a critical concern, as the risk of bridge failures poses significant economic and safety threats to the public. Traditional models often struggle to accurately estimate the equilibrium scour depth (deq) at bridge piers due to the complexity of scour processes and their reliance on simple regression methods. This study combines two metaheuristic optimization techniques—Siberian tiger optimization (STO) and brown-bear optimization algorithms (BOA)—with artificial neural networks (ANNs) to enhance deq prediction accuracy for both round- and sharp-nosed piers using both field and laboratory data. The findings indicate that BOA and STO effectively optimize ANN hyperparameters, resulting in improved prediction accuracy. Furthermore, both BOA–ANN and STO–ANN outperformed empirical equations and other machine learning techniques. An explicit equation was also derived from the BOA–ANN model. The influence of independent variables on deq was further examined using SHAP, revealing that pier width has the greatest impact on deq estimates.

AbstractIn this study, the generalized linear model (GLM) and four ensemble methods (partial least squares (PLS), boosting, bagging, and Bayesian) were applied to predict forest fire hazard in the Chalus Rood watershed in the Mazandaran Province, Iran. Data from 108 historical forest fire events collected through field surveys were applied as the basis of the analysis. About 70% of the data were used for training the models, while the remaining 30% was used for testing. A total of 14 environmental, climatic, and vegetation variables were used as input features to the models to predict forest fire probability. After conducting a multicollinearity test on the independent variables, the GLM and the ensemble models were applied for modeling. The efficiency of the models was evaluated using receiver operating characteristic (ROC) curve parameters. Results from the validation process, based on the area under the ROC curve (AUC), showed that the GLM, PLS-GLM, boosted-GLM, Bagging-GLM, and Bayesian-GLM models had efficiencies of 0.79, 0.75, 0.81, 0.84, and 0.85, respectively. The results indicated that all ensemble methods, except the PLS algorithm, improved the performance of the GLM model in modeling forest fire hazards in the Chalus Rood watershed, with the Bayesian algorithm being the most efficient method among them.

Abstract 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.

Gully erosion is one of the advanced forms of water erosion. Identifying the effective factors and gully erosion predicting is one of the important tools to control and manage such phenomenon. The main purpose of this study is to evaluate the effect of four different resampling algorithms including cross-validation (5-fold and 10-fold) and bootstrapping (Bootstrap and Optimism bootstrap) on boosted regression tree (BRT), support vector machine (SVM), and random forest (RF) models in spatial modeling and evaluation of head-cut gully erosion in Konduran watershed. For this purpose, based on an extensive field survey, the points of the head-cut of the gully erosion were identified first, and a map of the distribution of head-cut gully erosion in the study area was prepared. Then 18 variable identify and prepare as factors affecting the occurrence of head-cut gully erosion. To assess the efficiency of the models, receiver operating characteristics (ROC) and area under the curve (AUC) were used. The results of the assessment indicated that the use of resampling algorithms increases the efficiency of the models. The integrated optimism-bootstrap-BRT, optimism-bootstrap-SVM, and Optimism-Bootstrap-RF models with AUC 0.85, 0.823 and 0.89 respectively, outperformed the cross-validation 5fold (BRT, SVM, RF), Cross-validation 10fold (BRT, SVM, RF) and Bootstrap (BRT, SVM, RF) integrated algorithms.

Study region: Slovakia – 2927 municipalities. Study focus: This study maps and assesses the riverine flood potential in municipalities of Slovakia using the vector-based spatial multi-criteria analysis and geographic information systems (GIS). The flood potential index (FPI) in municipalities was computed based on eight flood factors, which were normalized, weighted and aggregated into the FPI index. Furthermore, we analyzed the spatial clusters of riverine flood potential using the local Moran’s coefficient of spatial autocorrelation. New hydrological insights for the region: The very high and high FPI (767 municipalities) was recorded mostly in municipalities of northern and eastern Slovakia. Very low and low FPI (1452 municipalities) dominates in western and southern Slovakia while moderate FPI (708 municipalities) was mainly recorded in municipalities of central Slovakia. The number of statistically significant high-high and low-low spatial clusters of riverine flood potential was 16 (524 municipalities) and 18 (487 municipalities), respectively. Clusters of positive spatial autocorrelation are mostly located in the northern and eastern Slovakia, one cluster in western Slovakia and seven clusters in central Slovakia. Clusters of negative spatial autocorrelation are mostly located in western and southern Slovakia and three clusters in eastern Slovakia. The proposed methodology can be useful for regular updating of Preliminary Flood Risk Assessment (PFRA) in Slovakia under the EU Floods Directive.

Floods are a consequence of extreme rainfall events. Although surface runoff generation is the origin of discharge, flood research usually focuses on lowlands where the impact is higher. Runoff and sediment delivery at slope and pedon scale receiving much less attention in the effort to understand flood behaviour in time and space. This is especially relevant in areas where, due to climatic and hydrogeological conditions, streams are ephemeral, so-called dry rivers (“wadis”, "ramblas" or “barrancos”) that are widespread throughout the Mediterranean. This paper researches the relationship between water delivery at pedon and slope scale with dry river floods in Macizo del Caroig, Eastern Iberian Peninsula. Plots of 1x1, 1x2, 1x4, and 2x8 m located in the “El Teularet” Soil Erosion and Degradation Research Station were monitored from 2004 to 2014 to measure soil and water delivery. Rainfall and flow at the dry river Barranco de Benacancil were also monitored. Results show that runoff and sediment discharge were concentrated in few events during the 11 years of research. A single flood event was registered in the channel on September 28, 2009, however, the runoff was registered 160 times at the plots. Runoff discharge was dependent on the size of the plots, with larger plots yielding lower runoff discharge per unit area, suggesting short runoff-travel distance and duration. Three rainfall events contributed with 26% of the whole runoff discharge, and five achieved 56% of the runoff. We conclude that the runoff generated at the plot scale is disconnected from the main channel. From a spatial point of view, there is a decrease in runoff coefficient along the slope. From a temporal point of view, the runoff is concentrated in a few rainfall events. These results show that the runoff generated at plot and slope scale does not contribute to the floods except for rainfall events with more than 100 mm day-1. The disconnection of the runoff and sediment delivery is confirmed by the reduction in the runoff delivery at plot scale due to the control of the length of the plot (slope) on the runoff and sediment delivery.

The purpose of the present study was to predict the areas affected by flood hazard in the Talar watershed, Mazandaran province, Iran, using Adaptive Boosting (AdaBoost), Boosted Generalized Linear Models (BGLM), Extreme Gradient Boosting (XGB) ensemble models, and the novel ensemble framework of deep decision trees include the Deep Boosting (DB) model. For this purpose, 14 flood conditioning variables were used as independent variables in flood hazard modeling. In addition, 130 flood points in the region were identified by field visits and available flood information, which were used as the dependent variable in modeling. The results showed that all used models have a good efficiency in predicting flood hazard. The area under curve (AUC) of BGLM, XGB, AdaBoost and DB models were 0.88, 0.87, 0.89 and 0.91, respectively, which indicated the highest efficiency of the DB model in flood hazard modeling in the study area. Relative importance of the variables showed that they have different effects in each model. Altitude and distance from the river are more important than other variables. However, these two variables have been selected as the most important variables based on machine learning models, but other variables may be influential in flood hazards.

The objective of this research is to propose and confirm a new machine learning approach of Best-First tree (BFtree), AdaBoost (AB), MultiBoosting (MB), and Bagging (Bag) ensembles for potential groundwater mapping and assessing role of influencing factors. The Yasuj-Dena area (Iran) is selected as a case study. For this regard, a Yasuj-Dena database was established with 362 springs locations and 12 groundwater-influencing factors (slope, aspect, elevation, stream power index (SPI), length of slope (LS), topographic wetness index (TWI), topographic position index (TPI), land use, lithology, distance from fault, distance from river, and rainfall). The database was employed to train and validate the proposed groundwater models. The area under the curve (AUC) and statistical metrics were employed to check and confirm the quality of the models. The result shows that the BFTree-Bag model (AUC = 0.810, kappa = 0.495) has the highest prediction performance, followed by the BFTree-MB model (AUC = 0.785, kappa = 0.477), and the BFTree-MB model (AUC = 0.745, kappa = 0.422). Compared to the benchmark of Random Forests, the BFTree-Bag model performs better; therefore, we conclude that the BFtree-Bag is a new tool should be used for modeling of groundwater potential.

Flooding is a natural disaster that causes considerable damage to different sectors and severely affects economic and social activities. The city of Saqqez in Iran is susceptible to flooding due to its specific environmental characteristics. Therefore, susceptibility and vulnerability mapping are essential for comprehensive management to reduce the harmful effects of flooding. The primary purpose of this study is to combine the Analytic Network Process (ANP) decision-making method and the statistical models of Frequency Ratio (FR), Evidential Belief Function (EBF), and Ordered Weight Average (OWA) for flood susceptibility mapping in Saqqez City in Kurdistan Province, Iran. The frequency ratio method was used instead of expert opinions to weight the criteria in the ANP. The ten factors influencing flood susceptibility in the study area are slope, rainfall, slope length, topographic wetness index, slope aspect, altitude, curvature, distance from river, geology, and land use/land cover. We identified 42 flood points in the area, 70% of which was used for modelling, and the remaining 30% was used to validate the models. The Receiver Operating Characteristic (ROC) curve was used to evaluate the results. The area under the curve obtained from the ROC curve indicates a superior performance of the ANP and EBF hybrid model (ANP-EBF) with 95.1% efficiency compared to the combination of ANP and FR (ANP-FR) with 91% and ANP and OWA (ANP-OWA) with 89.6% efficiency.

Gully erosion is the most active hydro-geomorphological phenomenon in the continental areas due to the high erosion rates triggered by the gully system. Monitoring and modelling gully development and gully distribution will contribute to understand landforms evolution and risk assessment. The purpose of the current research is to model head-cut gully erosion susceptibility (HCGES) using support vector machine (SVM), random forest (RF) and novel ensemble model of random forest with four Oblique methods (Logistic Regression, Ridge Regression, Partial least squares (PLS) and Support vector machine (SVM)) (hereafter called ensemble ORF) data mining models in Fareghan watershed, Hormozghan province, Iran. For this purpose, 14 variables influencing the gully development, were prepared and 145 head-cut gully erosion locations were identified in the study area. The efficiency of SVM, RF, ensemble ORF were evaluated based on receiver operating characteristic (ROC), the results have shown that all these three models are highly accurate and robust in predicting the head-cut gully erosion susceptibility zones. The results of the models were evaluated based on the area under the receiver operatic characteristic curve (AUC) in the validation stage presented that the efficiency of these models are 0.91, 0.94, and 0.96, respectively. Altitude and distance from the road in all three models were more important than other variables. The findings of this research will contribute to develop gully control strategies and to prevent the gully initiation where gully erosion is more susceptible.

Aflaj (plural of falaj) are tunnels or trenches built to deliver groundwater from its source to the point of consumption. Support vector machine (SVM) and extreme gradient boosting (XGB) machine learning models were used to predict groundwater aflaj potential in the Nizwa watershed in the Sultanate of Oman (Oman). Nizwa city is a focal point of aflaj that underlies the historical relationship between ecology, economic dynamics, agricultural systems, and human settlements. Three hyperparameter algorithms, grid search (GS), random search (RS), and Bayesian optimisation, were used to optimise the parameters of the XGB model. Sentinel-2 and light detection and ranging (LiDAR) data via geographical information systems (GIS) were employed to derive variables of land use/land cover, and hydrological, topographical, and geological factors. The groundwater aflaj potential maps were categorised into five classes: deficient, low, moderate, high, and very high. Based on the evaluation of accuracy in the training stage, the following models showed a high level of accuracy based on the area under the curve: Bayesian-XGB (0.99), GS-XGB (0.97), RS-XGB (0.96), SVM (0.96), and XGB (0.93). The validation results showed that the Bayesian hyperparameter algorithm significantly increased XGB model efficiency in modelling groundwater aflaj potential. The highest percentages of groundwater potential in the very high class were the XGB (10%), SVM (8%), GS-XGB (6%), RS-XGB (6%), and Bayesian-XGB (6%) models. Most of these areas were located in the central and northeast parts of the case study area. The study concluded that evaluating existing groundwater datasets, facilities, current, and future spatial datasets is critical in order to design systems capable of mapping groundwater aflaj based on geospatial and ML techniques. In turn, groundwater protection service projects and integrated water source management (IWSM) programs will be able to protect the aflaj irrigation system from threats by implementing timely preventative measures.

Large damages and losses resulting from floods are widely reported across the globe. Thus, the identification of the flood-prone zones on a flood susceptibility map is very essential. To do so, 13 conditioning factors influencing the flood occurrence in Brisbane river catchment in Australia (i.e., topographic, water-related, geological, and land use factors) were acquired for further processing and modeling. In this study, artificial neural networks (ANN), deep learning neural networks (DLNN), and optimized DLNN using particle swarm optimization (PSO) were exploited to predict and estimate the susceptible areas to the future floods. The significance of the conditioning factors analysis for the region highlighted that altitude, distance from river, sediment transport index (STI), and slope played the most important roles, whereas stream power index (SPI) did not contribute to the hazardous situation. The performance of the models was evaluated against the statistical tests such as sensitivity, specificity, the area under curve (AUC), and true skill statistic (TSS). DLNN and PSO-DLNN models obtained the highest values of sensitivity (0.99) for the training stage to compare with ANN. Moreover, the validations of specificity and TSS for PSO-DLNN recorded the highest values of 0.98 and 0.90, respectively, compared with those obtained by ANN and DLNN. The best accuracies by AUC were evaluated in PSO-DLNN (0.99 in training and 0.98 in testing datasets), followed by DLNN and ANN. Therefore, the optimized PSO-DLNN proved its robustness to compare with other methods.

This study predicts forest fire susceptibility in Chaloos Rood watershed in Iran using three machine learning (ML) models—multivariate adaptive regression splines (MARS), support vector machine (SVM), and boosted regression tree (BRT). The study utilizes 14 set of fire predictors derived from vegetation indices, climatic variables, environmental factors, and topographical features. To assess the suitability of the models and estimating the variance and bias of estimation, the training dataset obtained from the Natural Resources Directorate of Mazandaran province was subjected to resampling using cross validation (CV), bootstrap, and optimism bootstrap techniques. Using variance inflation factor (VIF), weight indicating the strength of the spatial relationship of the predictors to fire occurrence was assigned to each contributing variable. Subsequently, the models were trained and validated using the receiver operating characteristics (ROC) area under the curve (AUC) curve. Results of the model validation based on the resampling techniques (non, 5- and 10-fold CV, bootstrap and optimism bootstrap) produced AUC values of 0.78, 0.88, 0.90, 0.86 and 0.83 for the MARS model; 0.82, 0.82, 0.89, 0.87, 0.84 for the SVM and 0.87, 0.90, 0.90, 0.90, 0.91 for the BRT model. Across the individual model, the 10-fold CV performed best in MARS and SVM with AUC values of 0.90 and 0.89. Overall, the BRT outperformed the other models in all ramification with highest AUC value of 0.91 using optimism bootstrap resampling algorithm. Generally, the resampling process enhanced the prediction performance of all the models.

Flash flooding is considered one of the most dynamic natural disasters for which measures need to be taken to minimize economic damages, adverse effects, and consequences by mapping flood susceptibility. Identifying areas prone to flash flooding is a crucial step in flash flood hazard management. In the present study, the Kalvan watershed in Markazi Province, Iran, was chosen to evaluate the flash flood susceptibility modeling. Thus, to detect flash flood-prone zones in this study area, five machine learning (ML) algorithms were tested. These included boosted regression tree (BRT), random forest (RF), parallel random forest (PRF), regularized random forest (RRF), and extremely randomized trees (ERT). Fifteen climatic and geo-environmental variables were used as inputs of the flash flood susceptibility models. The results showed that ERT was the most optimal model with an area under curve (AUC) value of 0.82. The rest of the models’ AUC values, i.e., RRF, PRF, RF, and BRT, were 0.80, 0.79, 0.78, and 0.75, respectively. In the ERT model, the areal coverage for very high to moderate flash flood susceptible area was 582.56 km2 (28.33%), and the rest of the portion was associated with very low to low susceptibility zones. It is concluded that topographical and hydrological parameters, e.g., altitude, slope, rainfall, and the river’s distance, were the most effective parameters. The results of this study will play a vital role in the planning and implementation of flood mitigation strategies in the region.

Soil erosion determines landforms, soil formation and distribution, soil fertility, and land degradation processes. In arid and semiarid ecosystems, soil erosion is a key process to understand, foresee, and prevent desertification. Addressing soil erosion throughout watersheds scales requires basic information to develop soil erosion control strategies and to reduce land degradation. To assess and remediate the non-sustainable soil erosion rates, restoration programs benefit from the knowledge of the spatial distribution of the soil losses to develop maps of soil erosion. This study presents Support Vector Machine (SVM), Random Forest (RF), and adaptive boosting (AdaBoost) data mining models to map soil erosion susceptibility in Kozetopraghi watershed, Iran. A soil erosion inventory map was prepared from field rainfall simulation experiments on 174 randomly selected points along the Kozetopraghi watershed. In previous studies, this map has been prepared using indirect methods such as the Universal Soil Loss Equation to assess soil erosion. Direct field measurements for mapping soil erosion susceptibility have so far not been carried out in our study site in the past. The soil erosion rate data generated by simulated rainfall in 1 m2 plots at rainfall rate of 40 mmh−1 was used to develop the soil erosion map. Of the available data, 70% and 30% were randomly classified to calibrate and validate the models, respectively. As a result, the RF model with the highest area under the curve (AUC) value in a receiver operating characteristics (ROC) curve (0.91), and the lowest mean square error (MSE) value (0.09), has the most concordance and spatial differentiation. Sensitivity analysis by Jackknife and IncNodePurity methods indicates that the slope angle is the most important factor within the soil erosion susceptibility map. The RF susceptibility map showed that the areas located in the center and near the watershed outlet have the most susceptibility to soil erosion. This information can be used to support the development of sustainable restoration plans with more accuracy. Our methodology has been evaluated and can be also applied in other regions.

Prediction of the groundwater nitrate concentration is of utmost importance for pollution control and water resource management. This research aims to model the spatial groundwater nitrate concentration in the Marvdasht watershed, Iran, based on several artificial intelligence methods of support vector machine (SVM), Cubist, random forest (RF), and Bayesian artificial neural network (Baysia-ANN) machine learning models. For this purpose, 11 independent variables affecting groundwater nitrate changes include elevation, slope, plan curvature, profile curvature, rainfall, piezometric depth, distance from the river, distance from residential, Sodium (Na), Potassium (K), and topographic wetness index (TWI) in the study area were prepared. Nitrate levels were also measured in 67 wells and used as a dependent variable for modeling. Data were divided into two categories of training (70%) and testing (30%) for modeling. The evaluation criteria coefficient of determination (R2), mean absolute error (MAE), root mean square error (RMSE), and Nash–Sutcliffe efficiency (NSE) were used to evaluate the performance of the models used. The results of modeling the susceptibility of groundwater nitrate concentration showed that the RF (R2 = 0.89, RMSE = 4.24, NSE = 0.87) model is better than the other Cubist (R2 = 0.87, RMSE = 5.18, NSE = 0.81), SVM (R2 = 0.74, RMSE = 6.07, NSE = 0.74), Bayesian-ANN (R2 = 0.79, RMSE = 5.91, NSE = 0.75) models. The results of groundwater nitrate concentration zoning in the study area showed that the northern parts of the case study have the highest amount of nitrate, which is higher in these agricultural areas than in other areas. The most important cause of nitrate pollution in these areas is agriculture activities and the use of groundwater to irrigate these crops and the wells close to agricultural areas, which has led to the indiscriminate use of chemical fertilizers by irrigation or rainwater of these fertilizers is washed and penetrates groundwater and pollutes the aquifer.

The estimation and mapping of forest stand characteristics are vital because this information is necessary for sustainable forest management. The present study considers the use of a Bayesian additive regression trees (BART) algorithm as a non-parametric classifier using Sentinel-2A data and topographic variables to estimate the forest stand characteristics, namely the basal area (m2/ha), stem volume (m3/ha), and stem density (number/ha). These results were compared with those of three other popular machine learning (ML) algorithms, such as generalised linear model (GLM), K-nearest neighbours (KNN), and support vector machine (SVM). A feature selection was done on 28 variables including the multi-spectral bands on Sentinel-2 satellite, related vegetation indices, and ancillary data (elevation, slope, and topographic solar-radiation index derived from digital elevation model (DEM)) and then the most insignificant variables were removed from the datasets by recursive feature elimination (RFE). The study area was a mountainous forest with high biodiversity and an elevation gradient from 26 to 1636 m. An inventory dataset of 1200 sample plots was provided for training and testing the algorithms, and the predictors were fed into the ML models to compute and predict the forest stand characteristics. The accuracies and certainties of the ML models were assessed by their root mean square error (RMSE), mean absolute error (MAE), and R-squared (R2) values. The results demonstrated that BART generated the best basal area and stem volume predictions, followed by GLM, SVM, and KNN. The best RMSE values for both basal area (8.12 m2/ha) and stem volume (29.28 m3/ha) estimation were obtained by BART. Thus, the ability of the BART model for forestry application was established. On the other hand, KNN exhibited the highest RMSE values for all stand variable predictions, thereby exhibiting the least accuracy for this specific application. Moreover, the effectiveness of the narrow Sentinel-2 bands around the red edge and elevation was highlighted for predicting the forest stand characteristics. Therefore, we concluded that the combination of the Sentinel-2 products and topographic variables derived from the PALSAR data used in this study improved the estimation of the forest attributes in temperate forests.

Piping erosion is one form of water erosion that leads to significant changes in the landscape and environmental degradation. In the present study, we evaluated piping erosion modeling in the Zarandieh watershed of Markazi province in Iran based on random forest (RF), support vector machine (SVM), and Bayesian generalized linear models (Bayesian GLM) machine learning algorithms. For this goal, due to the importance of various geo-environmental and soil properties in the evolution and creation of piping erosion, 18 variables were considered for modeling the piping erosion susceptibility in the Zarandieh watershed. A total of 152 points of piping erosion were recognized in the study area that were divided into training (70%) and validation (30%) for modeling. The area under curve (AUC) was used to assess the effeciency of the RF, SVM, and Bayesian GLM. Piping erosion susceptibility results indicated that all three RF, SVM, and Bayesian GLM models had high efficiency in the testing step, such as the AUC shown with values of 0.9 for RF, 0.88 for SVM, and 0.87 for Bayesian GLM. Altitude, pH, and bulk density were the variables that had the greatest influence on the piping erosion susceptibility in the Zarandieh watershed. This result indicates that geo-environmental and soil chemical variables are accountable for the expansion of piping erosion in the Zarandieh watershed.

This study aims to evaluate a new approach in modeling gully erosion susceptibility (GES) based on a deep learning neural network (DLNN) model and an ensemble particle swarm optimization (PSO) algorithm with DLNN (PSO-DLNN), comparing these approaches with common artificial neural network (ANN) and support vector machine (SVM) models in Shirahan watershed, Iran. For this purpose, 13 independent variables affecting GES in the study area, namely, altitude, slope, aspect, plan curvature, profile curvature, drainage density, distance from a river, land use, soil, lithology, rainfall, stream power index (SPI), and topographic wetness index (TWI), were prepared. A total of 132 gully erosion locations were identified during field visits. To implement the proposed model, the dataset was divided into the two categories of training (70%) and testing (30%). The results indicate that the area under the curve (AUC) value from receiver operating characteristic (ROC) considering the testing datasets of PSO-DLNN is 0.89, which indicates superb accuracy. The rest of the models are associated with optimal accuracy and have similar results to the PSO-DLNN model; the AUC values from ROC of DLNN, SVM, and ANN for the testing datasets are 0.87, 0.85, and 0.84, respectively. The efficiency of the proposed model in terms of prediction of GES was increased. Therefore, it can be concluded that the DLNN model and its ensemble with the PSO algorithm can be used as a novel and practical method to predict gully erosion susceptibility, which can help planners and managers to manage and reduce the risk of this phenomenon.

In the present study, gully erosion susceptibility was evaluated for the area of the Robat Turk Watershed in Iran. The assessment of gully erosion susceptibility was performed using four state-of-the-art data mining techniques: random forest (RF), credal decision trees (CDTree), kernel logistic regression (KLR), and best-first decision tree (BFTree). To the best of our knowledge, the KLR and CDTree algorithms have been rarely applied to gully erosion modeling. In the first step, from the 242 gully erosion locations that were identified, 70% (170 gullies) were selected as the training dataset, and the other 30% (72 gullies) were considered for the result validation process. In the next step, twelve gully erosion conditioning factors, including topographic, geomorphological, environmental, and hydrologic factors, were selected to estimate gully erosion susceptibility. The area under the ROC curve (AUC) was used to estimate the performance of the models. The results revealed that the RF model had the best performance (AUC = 0.893), followed by the KLR (AUC = 0.825), the CDTree (AUC = 0.808), and the BFTree (AUC = 0.789) models. Overall, the RF model performed significantly better than the others, which may support the application of this method to a transferable susceptibility model in other areas. Therefore, we suggest using the RF, KLR, and CDT models for gully erosion susceptibility mapping in other prone areas to assess their reproducibility.

Gully erosion destroys agricultural and domestic grazing land in many countries, especially those with arid and semi-arid climates and easily eroded rocks and soils. It also generates large amounts of sediment that can adversely impact downstream river channels. The main objective of this research is to accurately detect and predict areas prone to gully erosion. In this paper, we couple hybrid models of a commonly used base classifier (reduced pruning error tree, REPTree) with AdaBoost (AB), bagging (Bag), and random subspace (RS) algorithms to create gully erosion susceptibility maps for a sub-basin of the Shoor River watershed in northwestern Iran. We compare the performance of these models in terms of their ability to predict gully erosion and discuss their potential use in other arid and semi-arid areas. Our database comprises 242 gully erosion locations, which we randomly divided into training and testing sets with a ratio of 70/30. Based on expert knowledge and analysis of aerial photographs and satellite images, we selected 12 conditioning factors for gully erosion. We used multi-collinearity statistical techniques in the modeling process, and checked model performance using statistical indexes including precision, recall, F-measure, Matthew correlation coefficient (MCC), receiver operatic characteristic curve (ROC), precision−recall graph (PRC), Kappa, root mean square error (RMSE), relative absolute error (PRSE), mean absolute error (MAE), and relative absolute error (RAE). Results show that rainfall, elevation, and river density are the most important factors for gully erosion susceptibility mapping in the study area. All three hybrid models that we tested significantly enhanced and improved the predictive power of REPTree (AUC=0.800), but the RS-REPTree (AUC= 0.860) ensemble model outperformed the Bag-REPTree (AUC= 0.841) and the AB-REPTree (AUC= 0.805) models. We suggest that decision makers, planners, and environmental engineers employ the RS-REPTree hybrid model to better manage gully erosion-prone areas in Iran.