Search

Nitrate pollution of groundwater has increased dramatically worldwide due to increase of population and agricultural productivity. The resulting nitrate concentration in groundwater is usually a combination of various types of point and non-point pollutant sources. It is often difficult to distinguish between these sources since groundwater is formed in large and complex catchments with various natural processes and anthropogenic influence that contribute to a certain downstream nitrate concentration. For such conditions, this paper uses a methodology that can be used to inversely determine type and location of main nitrate pollutant source. The methodology builds on two state-of-the-art data mining techniques, boosted regression tree (BRT) and k-nearest neighbor (KNN). These techniques are used to produce a nitrate pollution vulnerability map. The methodology can mitigate effects of subjective judgement on determining importance of different sources and mechanisms for nitrate transport. The investigated mechanisms are hydrogeological, hydrological, anthropogenic, topography, and soil conditioning factors. Thus, the proposed methodology is used to separate between natural processes and anthropogenic effects on nitrate pollution. To calculate the groundwater vulnerability maps, a groundwater nitrate concentration of 40 mg/L (suggested by WHO with a 20% risk margin) was selected as a general threshold for identifying polluted areas that resulted in 96 polluted wells. Non-polluted locations were selected from well data with nitrate concentration less than 15 mg/L (96 non-polluted). The models were trained on 70% polluted and 70% non-polluted site data. The remaining data, 30% polluted and 30% non-polluted sites, were used to validate the simulation results. Results showed that the BRT produced outputs with higher performance than the KNN algorithm. The final ranking results based on the BRT model showed the higher importance of hydraulic conductivity, river density, soil, slope percent, net recharge, and distance from villages, in order, relative to other factors. [ABSTRACT FROM AUTHOR]

Freshwater scarcity is an ever-increasing problem throughout the arid and semi-arid countries, and it often results in poverty. Thus, it is necessary to enhance understanding of freshwater resources availability, particularly for groundwater, and to be able to implement functional water resources plans. This study introduces a novel statistical approach combined with a data-mining ensemble model, through implementing evidential belief function and boosted regression tree (EBF-BRT) algorithms for groundwater potential mapping of the Lordegan aquifer in central Iran. To do so, spring locations are determined and partitioned into two groups for training and validating the individual and ensemble methods. In the next step, 12 groundwater-conditioning factors (GCFs), including topographical and hydrogeological factors, are prepared for the modeling process. The mentioned factors are employed in the application of the EBF model. Then, the EBF values of the GCFs are implemented as input to the BRT algorithm. The results of the modeling process are plotted to produce spring (groundwater) potential maps. To verify the results, the receiver operating characteristics (ROC) test is applied to the model's output. The findings of the test indicated that the areas under the ROC curves are 75 and 82% for the EBF and EBF-BRT models, respectively. Therefore, it can be inferred that the combination of the two techniques could increase the efficacy of these methods in groundwater potential mapping. [ABSTRACT FROM AUTHOR]

Graphical abstract Highlights • A novel hybrid model, namely LMT, was applied for groundwater potential mapping. • The result of LMT was validated using AUC-ROC method. • The accuracy of LMT was comprised with C4.5 and SVM models. • The capability and robustness of all three models were assessed. • Relative slope position (RSP) was the most important conditioning factor. Abstract Sustainable water resources management in arid and semi-arid areas needs robust models, which allow accurate and reliable predictive modeling. This issue has motivated the researchers to develop hybrid models that offer solutions on modelling problems and accurate predictions of groundwater potential zonation. For this purpose, this research aims to investigate the capability and robustness of a novel hybrid model, namely the logistic model tree (LMT) and compares it with state-of-the-art models such as the support vector machine and C4.5 models that locate potential zones for groundwater springs. A spring location dataset consisting of 359 springs was provided by field surveys and national reports and from which three different sample data sets (S1–S3) were randomly prepared (70% for training and 30% for validation). Additionally, 16 spring-related factors were analyzed using regression logistic analysis to find which factors play a significant role in spring occurrence. Twelve significant geo-environmental and morphometric factors were identified and applied in all models. The accuracy of models was evaluated by three different threshold-dependent and –Independent methods including efficiency (E), true skill statistic (TSS), and area under the receiver operating characteristics curve (AUC-ROC) methods. Results showed that the LMT model had the highest accuracy performance for all three validation datasets (E mean = 0.860, TSS mean = 0.718, AUC-ROC mean = 0.904); although a slight sensitivity to change in input data was sometimes observed for this model. Furthermore, the findings showed that relative slope position (RSP) was the most important factor followed by distance from faults and lithology. [ABSTRACT FROM AUTHOR]

Considering the unstable condition of water resources in Iran and many other countries in arid and semi-arid regions, groundwater studies are very important. Therefore, the aim of this study is to model groundwater potential by qanat locations as indicators and ten advanced and soft computing models applied to the Beheshtabad Watershed, Iran. Qanat is a man-made underground construction which gathers groundwater from higher altitudes and transmits it to low land areas where it can be used for different purposes. For this purpose, at first, the location of the qanats was detected using extensive field surveys. These qanats were classified into two datasets including training (70%) and validation (30%). Then, 14 influence factors depicting the region’s physical, morphological, lithological, and hydrological features were identified to model groundwater potential. Linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), flexible discriminant analysis (FDA), penalized discriminant analysis (PDA), boosted regression tree (BRT), random forest (RF), artificial neural network (ANN), K-nearest neighbor (KNN), multivariate adaptive regression splines (MARS), and support vector machine (SVM) models were applied in R scripts to produce groundwater potential maps. For evaluation of the performance accuracies of the developed models, ROC curve and kappa index were implemented. According to the results, RF had the best performance, followed by SVM and BRT models. Our results showed that qanat locations could be used as a good indicator for groundwater potential. Furthermore, altitude, slope, plan curvature, and profile curvature were found to be the most important influence factors. On the other hand, lithology, land use, and slope aspect were the least significant factors. The methodology in the current study could be used by land use and terrestrial planners and water resource managers to reduce the costs of groundwater resource discovery. [ABSTRACT FROM AUTHOR]

In recent years, application of ensemble models has been increased tremendously in various types of natural hazard assessment such as landslides and floods. However, application of this kind of robust models in groundwater potential mapping is relatively new. This study applied four data mining algorithms including AdaBoost, Bagging, generalized additive model (GAM), and Naive Bayes (NB) models to map groundwater potential. Then, a novel frequency ratio data mining ensemble model (FREM) was introduced and evaluated. For this purpose, eleven groundwater conditioning factors (GCFs), including altitude, slope aspect, slope angle, plan curvature, stream power index (SPI), river density, distance from rivers, topographic wetness index (TWI), land use, normalized difference vegetation index (NDVI), and lithology were mapped. About 281 well locations with high potential were selected. Wells were randomly partitioned into two classes for training the models (70% or 197) and validating them (30% or 84). AdaBoost, Bagging, GAM, and NB algorithms were employed to get groundwater potential maps (GPMs). The GPMs were categorized into potential classes using natural break method of classification scheme. In the next stage, frequency ratio (FR) value was calculated for the output of the four aforementioned models and were summed, and finally a GPM was produced using FREM. For validating the models, area under receiver operating characteristics (ROC) curve was calculated. The ROC curve for prediction dataset was 94.8, 93.5, 92.6, 92.0, and 84.4% for FREM, Bagging, AdaBoost, GAM, and NB models, respectively. The results indicated that FREM had the best performance among all the models. The better performance of the FREM model could be related to reduction of over fitting and possible errors. Other models such as AdaBoost, Bagging, GAM, and NB also produced acceptable performance in groundwater modelling. The GPMs produced in the current study may facilitate groundwater exploitation by determining high and very high groundwater potential zones. [ABSTRACT FROM AUTHOR]

• Integration of InSAR and machine learning to study ground deformation. • Ground deformation modelling over agricultural areas. • Relating topographical, hydrogeological, and anthropogenic factors to subsidence. • High performance of the BRT and XGB algorithms in modelling ground deformation. • Pumping wells, faults, altitude, and water table had significant contribution. Land subsidence is an increasing human-induced disaster that not only damages building and transportation structures but also diminishes the water storage capacity of the aquifers. Land subsidence is a very complex phenomenon impacted by various geo-environmental and hydrological factors. Application of the interferometric synthetic aperture radar (InSAR) is becoming a common approach to detect land subsidence rates, though, it suffers from the lack of continuity over the spatial surfaces due to the vegetation decorrelation, coverage alterations (cultivation and non-cultivation seasons), in the agricultural areas, and rough topography. The lack of continuity can, however, be resolved using artificial intelligence. In our case study, while InSAR deformation data only covered ∼ 2% of the plain's surface, we employed boosted regression trees (BRT) and extreme gradient boosting (XGB) algorithms to provide a full coverage map of the groundwater-induced land subsidence based on the InSAR analysis. For this, a set of topographical, hydrological, hydrogeological, and anthropogenic factors was selected. The InSAR and input factors' resolution data were resampled to a 100-by-100 m to match. The implemented models predicted the long-term deformation rate with the acceptable performances of the BRT (RMSE = 3.3 mm/year, MAE = 2.0 mm/year, R2 = 0.985) and the XGB with linear booster (RMSE = 3.5 mm/year, MAE = 2.1 mm/year, R2 = 0.983). Considering the substantial ground deformation in the studied area (from −216 to 49 mm/year), RMSE values of 3.3, and 3.5 mm/year between the InSAR measurement and model predictions show great potential for combined InSAR-machine learning technique for pumping-driven land subsidence studies. Thus, the introduced approach is suggested for other areas being damaged by excessive pumping and agricultural development to produce an accurate full coverage map of subsidence. [ABSTRACT FROM AUTHOR]

In recent years, assessment of sediment contamination by heavy metals, i.e., arsenic, has attracted the interest of scientists worldwide. The present study provides a new methodology to better understand the factors influencing surface water vulnerability to arsenic pollution by two advanced machine learning algorithms including boosted regression trees (BRT) and random forest (RF). Based on the sediment quality guidelines (Effects range low) polluted and non-polluted arsenic sediment samples were defined with concentrations >8 ppm and <8 ppm, respectively. Different conditioning factors such as topographical, lithology, erosion, hydrological, and anthropogenic factors were acquired to model surface waters' vulnerability to arsenic. We trained and validated the models using 70 and 30% of both polluted and non-polluted samples, respectively, and generated surface vulnerability maps. To verify the maps to arsenic pollution, the receiver operating characteristics (ROC) curve was implemented. The results approved the acceptable performance of the RF and BRT algorithms with an area under ROC values of 85% and 75.6%, respectively. Further, the findings showed higher importance of precipitation, slope aspect, distance from residential areas, and slope length in arsenic pollution in the modeling process. Erosion, lithology, and land use maps were introduced as the least important factors. The introduced methodology can be used to define the most vulnerable areas to arsenic pollution in advance and implement proper remediation actions to reduce the damages. • About 27% of the riparian zone has very high vulnerability to arsenic pollution. • RF had area under ROC curve of 85% in spatial modeling of arsenic at rivers. • About 70% polluted areas are located in the vicinity of residential areas and mining activities. • Precipitation, slope aspect, distance from residential areas were the most important factor in the modeling process. [ABSTRACT FROM AUTHOR]

In this study, a new hybridized machine learning algorithm for urban flood susceptibility mapping, named MultiB-MLPNN, was developed using a multi-boosting technique and MLPNN. The model was tested in Amol City, Iran, a data-scarce city in an ungauged area which is prone to severe flood inundation events and currently lacks flood prevention infrastructure. Performance of the hybridized model was compared with that of a standalone MLPNN model, random forest and boosted regression trees. Area under the curve, efficiency, true skill statistic, Matthews correlation coefficient, misclassification rate, sensitivity and specificity were used to evaluate model performance. In validation, the MultiB-MLPNN model showed the best predictive performance. The hybridized MultiB-MLPNN model is thus useful for generating realistic flood susceptibility maps for data-scarce urban areas. The maps can be used to develop risk-reduction measures to protect urban areas from devastating floods, particularly where available data are insufficient to support physically based hydrological or hydraulic models. [ABSTRACT FROM AUTHOR]

• Spotting high potential areas for groundwater exploitation. • Application of two state-of-the-art machine learning algorithms. • Feeding only DEM-derived factors to the algorithms to determine groundwater potential. • Extreme gradient boosting and parallel random forest yielded AUC values of about 86% • Convergence, TWI, river density, and altitude are the most important factors. Groundwater (GW) resources provide a large share of the world's water demand for various sections such as agriculture, industry, and drinking water. Particularly in the arid and semi-arid regions, with surface water scarcity and high evaporation, GW is a valuable commodity. Yet, GW data are often incomplete or nonexistent. Therefore, it is a challenge to achieve a GW potential assessment. In this study, we developed methods to produce reliable GW potential maps (GWPM) with only digital elevation model (DEM)-derived data as inputs. To achieve this objective, a case study area in Iran was selected and 13 factors were extracted from the DEM. A spring location dataset was obtained from the water sector organizations and, along with the non-spring locations, fed into machine learning algorithms for training and validation. For delineating reliable GW potential, algorithms including random forest (RF) and its developed version, parallel RF (PRF), as well as extreme gradient boosting (XGB) with different boosters were used. The area under the receiver operating characteristics curve indicated that the PRF and XGB with linear booster give similar high accuracy (about 86%) for GWPM. The most important factors for accurate GWPM in the modeling procedure were convergence, topographic wetness index, river density, and altitude. Overall, we conclude that high-accuracy GWPMs can be produced with only DEM-derived factors with acceptable accuracy. The developed methodology can be employed to produce initial information for GW exploitation in areas facing a lack of data. [ABSTRACT FROM AUTHOR]

Groundwater (GW) is being uncontrollably exploited in various parts of the world resulting from huge needs for water supply as an outcome of population growth and industrialization. Bearing in mind the importance of GW potential assessment in reaching sustainability, this study seeks to use remote sensing (RS)-derived driving factors as an input of the advanced machine learning algorithms (MLAs), comprising deep boosting and logistic model trees to evaluate their efficiency. To do so, their results are compared with three benchmark MLAs such as boosted regression trees, k-nearest neighbors, and random forest. For this purpose, we firstly assembled different topographical, hydrological, RS-based, and lithological driving factors such as altitude, slope degree, aspect, slope length, plan curvature, profile curvature, relative slope position, distance from rivers, river density, topographic wetness index, land use/land cover (LULC), normalized difference vegetation index (NDVI), distance from lineament, lineament density, and lithology. The GW spring indicator was divided into two classes for training (434 springs) and validation (186 springs) with a proportion of 70:30. The training dataset of the springs accompanied by the driving factors were incorporated into the MLAs and the outputs were validated by different indices such as accuracy, kappa, receiver operating characteristics (ROC) curve, specificity, and sensitivity. Based upon the area under the ROC curve, the logistic model tree (87.813%) generated similar performance to deep boosting (87.807%), followed by boosted regression trees (87.397%), random forest (86.466%), and k-nearest neighbors (76.708%) MLAs. The findings confirm the great performance of the logistic model tree and deep boosting algorithms in modelling GW potential. Thus, their application can be suggested for other areas to obtain an insight about GW-related barriers toward sustainability. Further, the outcome based on the logistic model tree algorithm depicts the high impact of the RS-based factor, such as NDVI with 100 relative influence, as well as high influence of the distance from river, altitude, and RSP variables with 46.07, 43.47, and 37.20 relative influence, respectively, on GW potential. [ABSTRACT FROM AUTHOR]

Land subsidence (LS) is among the most critical environmental problems, affecting both agricultural sustainability and urban infrastructure. Existing methods often use either simple regression models or complex hydraulic models to explain and predict LS. There are few studies that identify the risk factors and predict the risk of LS using machine learning models. This study compares four tree-based machine learning models for land subsidence hazard modelling at a study area in Hamadan plain (Iran). The study also analyzes the importance of six risk factors including topography (elevation, slope), geomorphology (distance from stream, drainage density), hydrology (groundwater drawdown) and lithology on LS. Thematic layers of each variable related to the LS phenomenon are prepared and utilized as the inputs to the four tree-based machine learning models, including the Rule-Based Decision Tree (RBDT), Boosted Regression Trees (BRT), Classification And Regression Tree (CART), and the Random Forest (RF) algorithms to produce a consolidated LS hazard map. The accuracy of the generated maps is then evaluated using the area under the receiver operating characteristic curve (AUC) and the True Skill Statistics (TSS). The RF approach had the lowest predictive error for mapping the LS hazard (i.e., AUC 96.7% for training, AUC 93.8% for validation, TSS 0.912 for training, TSS 0.904 for validation) followed by BRT. Groundwater drawdown was seen to be the most influential factor that contributed to land subsidence in the present study area, followed by lithology and distance from the stream network. Unlabelled Image • Four tree-based models were applied for land subsidence modelling. • RF is the best model to spatially predict land subsidence hazard. • Groundwater drawdown was the most influential factor. [ABSTRACT FROM AUTHOR]

The main aim of this study was to apply and compare two GIS-based data mining models, namely support vector machine (SVM) by four kernel functions (linear-SVM, polynomial-SVM, radial basic function-SVM, and sigmoidal-SVM) and entropy models in landslide susceptibility mapping, in Shangzhou District, China. Initially, 145 landslide locations were mapped using early reports, aerial photographs, and supported by field surveys. Subsequently, landslides in the study area were divided randomly into training and validation datasets (70/30) using ArcGIS 10.0 software. In the current study, 14 landslide conditioning factors, namely, slope aspect, slope angle, profile curvature, plan curvature, altitude, topographic wetness index (TWI), stream power index (SPI), sediment transport index (STI), normalized difference vegetation index (NDVI), distance from roads, distance from rivers, distance from faults, rainfall, and lithology, were exploited to detect the most susceptible areas. In the next step, landslide susceptibility maps generated by four types of SVM or entropy models were produced. Finally, validation of the landslide susceptibility maps produced by different models was evaluated using receiver operating characteristics (ROC) curves. The results showed that the entropy model exhibited the highest success rate (0.7610), followed by polynomial-SVM (0.7526), the sigmoidal-SVM (0.7518), radial basic function-SVM (0.7446), and linear-SVM (0.7390) models. Similarly, the ROC plots also showed that the prediction rates gave almost similar results. The entropy model had the highest prediction rate (0.7599), followed by polynomial-SVM (0.7259), sigmoidal-SVM (0.7203), radial basic function-SVM (0.7149), and linear-SVM (0.7009) models. Hence, it can be concluded that the five models used in this study gave close results, with the entropy model exhibiting best performance in landslide susceptibility mapping. [ABSTRACT FROM AUTHOR]

The main objective of the current study is to apply a random forest (RF) data-driven model and prioritization of landslide conditioning factors according to this method and its comparison to a multivariate adaptive regression spline (MARS) model for landslide susceptibility mapping in China. For this purpose, at first, landslide locations were identified by earlier reports, aerial photographs, and field surveys and a total of 348 landslides were mapped from various sources in GIS. Then, the landslide inventory was randomly split into a training dataset (70% = 244 landslides) and the remaining (30% = 104 landslides) were used for validation. In this study, 12 landslide conditioning factors were applied to detect the most susceptible areas. These factors were slope aspect, altitude, distance to faults, lithology, normalized difference vegetation index, plan curvature, profile curvature, distance to rivers, distance to roads, slope angle, stream power index, and topographic wetness index. The relationship between each conditioning factor and landslide was finalized using a frequency ration (FR) model. Subsequently, landslide-susceptible areas were mapped using the MARS and RF models. The results revealed that the most important conditioning factors according to the accuracy measure (mean decrease) of the RF model are lithology (23.47%), distance to faults (22.21%), and altitude (19.58%). We also notice that altitude (19.04%), distance to faults (18.83%), and distance to roads (15.29%) have the highest importance according to the Gini measure. Finally, the accuracy of the landslide susceptibility maps produced from the two models was verified using a receiver operating characteristics curve. The results showed that the landslide susceptibility map produced using the MARS model has a higher prediction rate than RF by area under the curve values of 87.51 and 77.32%, respectively. According to the validation results, the map produced by the MARS model exhibits the better accuracy and could be proposed for land-use planning in the study area. [ABSTRACT FROM AUTHOR]

Landslide is a natural hazard that results in many economic damages and human losses every year. Numerous researchers have studied landslide susceptibility mapping (LSM), each attempting to improve the accuracy of the final outputs. However, few studies have been published on the training data selection effects on the LSM. Thus, this study assesses the training landslides random selection effects on support vector machine (SVM) accuracy, logistic regression (LR) and artificial neural networks (ANN) models for LSM in a catchment at the Dodangeh watershed, Mazandaran province, Iran. A 160 landslide locations inventory was collected by Geological Survey of Iran for this investigation. Different methods were implemented to define the landslide locations, such as inventory reports, satellite images and field survey. Moreover, 14 landslide conditioning factors were considered in the analysis of landslide susceptibility. These factors include curvature, plan curvature, profile curvature, altitude, slope angle, slope aspect, distance to faults, distance to stream, topographic wetness index, stream power index, terrain roughness index, sediment transport index, lithology and land use. The results show that the random landslide training data selection affected the parameter estimations of the SVM, LR and ANN algorithms. The results also show that the training samples selection had an effect on the accuracy of the susceptibility model because landslide conditioning factors vary according to the geographic locations in the study area. The LR model was found to be less sensitive than the SVM and ANN models to the training samples selection. Validation results showed that SVM and LR models outperformed the ANN model for all scenarios. The average overall accuracy of LR, SVM and ANN models are 81.42%, 79.82% and 70.2%, respectively. [ABSTRACT FROM AUTHOR]

Most part of Iran is arid and semi-arid; thus in most parts of the region, groundwater is the only source of water. This research presents a method based on a spatial multi-criterion evaluation (SMCE) for designing possible sites of underground dams and ranks them according to their suitability. The method was tested for siting underground dams in the Alborz Province, Iran. At first, screening algorithm was applied using exclusionary criteria, and thirty-one potential areas were recognized in the study area. In the next step, a suitable gorge or valley was recognized using the combination of basic maps and extensive field surveys (long axis of tank level) in each potential area. Subsequently, the analytical hierarchy process was used as a powerful tool for decision-making in the SMCE in order to evaluate different criteria for underground dam sites. SMCE techniques were then applied to combine the criteria, and obtain a suitability map in the study area. These sites were then compared and ranked according to their main criteria such as water, storage, axis and socio-economics. All these criteria were assessed through geographical information system modelling. This method shows passable results and could be used for site selection of underground dams in other regions of Iran. [ABSTRACT FROM PUBLISHER]

Floods in urban environments often result in loss of life and destruction of property, with many negative socio-economic effects. However, the application of most flood prediction models still remains challenging due to data scarcity. This creates a need to develop novel hybridized models based on historical urban flood events, using, e.g., metaheuristic optimization algorithms and wavelet analysis. The hybridized models examined in this study (Wavelet-SVR-Bat and Wavelet-SVR-GWO), designed as intelligent systems, consist of a support vector regression (SVR), integrated with a combination of wavelet transform and metaheuristic optimization algorithms, including the grey wolf optimizer (GWO), and the bat optimizer (Bat). The efficiency of the novel hybridized and standalone SVR models for spatial modeling of urban flood inundation was evaluated using different cutoff-dependent and cutoff-independent evaluation criteria, including area under the receiver operating characteristic curve (AUC), Accuracy (A), Matthews Correlation Coefficient (MCC), Misclassification Rate (MR), and F-score. The results demonstrated that both hybridized models had very high performance (Wavelet-SVR-GWO: AUC = 0.981, A = 0.92, MCC = 0.86, MR = 0.07; Wavelet-SVR-Bat: AUC = 0.972, A = 0.88, MCC = 0.76, MR = 0.11) compared with the standalone SVR (AUC = 0.917, A = 0.85, MCC = 0.7, MR = 0.15). Therefore, these hybridized models are a promising, cost-effective method for spatial modeling of urban flood susceptibility and for providing in-depth insights to guide flood preparedness and emergency response services. [ABSTRACT FROM AUTHOR]

A quantitative understanding of the hydro-environmental factors that influence the occurrence of agricultural drought events would enable more strategic climate change adaptation and drought management plans. Practical drought hazard mapping remains challenging due to possible exclusion of the most pertinent drought drivers, and to the use of inadequate predictive models that cannot describe drought adequately. This research aims to develop new approaches to map agricultural drought hazard with state-of-the-art machine learning models, including classification and regression trees (CART), boosted regression trees (BRT), random forests (RF), multivariate adaptive regression splines (MARS), flexible discriminant analysis (FDA) and support vector machines (SVM). Hydro-environmental datasets were used to calculate the relative departure of soil moisture (RDSM) for eight severe droughts for drought-prone southeast Queensland, Australia, over the period 1994–2013. RDSM was then used to generate an agricultural drought inventory map. Eight hydro-environmental factors were used as potential predictors of drought. The goodness-of-fit and predictive performance of all models were evaluated using different threshold-dependent and threshold-independent methods, including the true skill statistic (TSS), Efficiency (E), F-score, and the area under the receiver operating characteristic curve (AUC-ROC). The RF model (AUC-ROC = 97.7%, TSS = 0.873, E = 0.929, F-score = 0.898) yielded the highest accuracy, while the FDA model (with AUC-ROC = 73.9%, TSS = 0.424, E = 0.719, F-score = 0.512) showed the worst performance. The plant available water holding capacity (PAWC), mean annual precipitation, and clay content were the most important variables to be used for predicting the agricultural drought. About 21.2% of the area is in high or very high drought risk classes, and therefore, warrant drought and environmental protection policies. Importantly, the models do not require data on the precipitation anomaly for any given drought year; the spatial patterns in AGH were consistent for all drought events, despite very different spatial patterns in precipitation anomaly among events. Such machine-learning approaches are able to construct an overall risk map, thus assisting in the adoption of a robust drought contingency planning measure not only for this area, but also, in other regions where drought presents a pressing challenge, including its influence on key practical dimensions of social, environmental and economic sustainability. Unlabelled Image • Spatial and temporal machine learning approaches proposed for drought risk mapping. • Performance of six machine learning models is compared. • MARS and RF are suitable machine learning models for spatial risk mapping • Approximately 26% of the study area is at high or very high drought risk. • New approach considers hydro-geo-environmental factors for drought risk policy. [ABSTRACT FROM AUTHOR]