Supervised Classification of Groundwater Potential Mapping Using Integrated Machine Learning and GIS-Based Techniques

Groundwater serves as a crucial resource in arid and semi-arid regions, as aquifers offer a sustainable source of high-quality water throughout the year when rainfall and surface water are scarce. Due to climate change, rainfall is anticipated to decrease, and droughts are expected to become more severe [1]. Generally, the term "groundwater potential" refers to the amount of groundwater available in an area and depends on various hydrologic and hydrogeological factors. From a hydrogeological perspective, this term can be defined as the probability of groundwater occurrence in an area or the prediction of where the highest borehole yields may be found. Proper assessment of groundwater potential can serve as a valuable guideline for decision-makers to identify appropriate groundwater strategies within an area and manage the aquifer system sustainably [2, 3].

Methods for mapping groundwater potential zones are classified into two categories: knowledge-driven techniques, such as the Analytical Hierarchy Process (AHP), which are based on expert experiences and thus influenced by specialist knowledge and subjectivity, and data-driven techniques, which involve probabilistic, statistical, and data mining approaches based on the empirical element of groundwater assessments [4-6]. Expert-based decision approaches have long been utilized, while machine learning is a comparatively newer method. A crucial difference between machine learning and expert techniques is that machine learning classification employs artificial intelligence capabilities to discover complex associations between explanatory factors that would otherwise remain unnoticed. Consequently, machine learning is well-suited for mapping complicated, spatially distributed variables such as groundwater occurrence [7]. Machine learning (ML) uses algorithms that can learn and improve themselves based on the experience they gain from collected data to produce accurate predictions [8]. ML can be divided into two categories: supervised and unsupervised algorithms. In unsupervised learning, the algorithm derives patterns from data that have not been labeled or otherwise categorized, aiming to understand the fundamental structure of the data. Since the model does not have a reference for the expected output data format, it cannot carry out any training on the model. Instead, the model investigates the structure of the data so that valuable information may be extracted from it [9]. Supervised learning algorithms are the preferred method when both input and output variables are clearly labeled. The goal is to learn patterns and correlations between variables based on prior experience (training data) and then apply that information to generate predictions on data that has not been seen or is unknown (test data). In supervised learning, two subcategories of data can be modeled: classification, in which a model aims to predict categorical or class labels, and regression, in which the models attempt to predict a continuous output. Both classification and regression are examples of predictive modeling [10]. The Groundwater Potential Mapping (GPM) literature presents several supervised classification strategies.

For instance, Ozdemir [11] employed a binary logistic regression to generate a groundwater spring potential map GPM of the Sultan Mountains in central Turkey using a logistic regression technique integrated with the Geographic Information System (GIS) environment. An area value of 0.82 was determined for the Receiver Operating Characteristic (ROC) curve model, indicating an accurate prediction of spring potential in the study region. Additionally, the existing groundwater spring test data and the created model were in close agreement. Nguyen et al. [12] utilized an advanced ensemble machine learning model (RABANN) that combines Artificial Neural Networks (ANN) with RealAdaBoost (RAB) ensemble approach to evaluate the groundwater potential of DakNong province, Vietnam. This work used twelve conditioning parameters and well-yield data to generate the training and testing datasets. The results indicate that the RAB ensemble approach improved the performance of the ANN model. With minor modifications to the input data, the ensemble-developed model could be applied to map the groundwater potential of other regions and countries to enhance water resource management.

Sarkar et al. [13] developed ensemble machine learning (EML) algorithms for groundwater potentiality mapping (GPM) in Bangladesh's Teesta River basin. These algorithms include random forests (RF) and random subspaces (RSS). The GPM was verified using the Receiver Operating Characteristics (ROC) curve. The random subspaces machine learning (RSS) model, with an AUC of 0.892, was determined to be the most accurate representation model for groundwater potentiality modeling, followed by random forests with an AUC of 0.86. The RSS model outperforms the RF mode in terms of groundwater potentiality models since a higher AUC indicates more accurate predictions of the model's output.

Phong et al. [14] proposed three unique ML hybrid models: modified RealAdaBoost (MRAB-FT), bagging (BA-FT), and rotation forest (RF-FT). All models are functional tree (FT) base classifiers capable of handling binary and multi-class target variables for GPM modeling in the DakLak region of Vietnam. This study employed a threshold value greater than 2 L/s for selecting training and testing datasets for groundwater classification as groundwater and non-groundwater.  Standard statistical methods assessed how well the created models performed (PPV, NPV, SST, SPC, Kappa, RMSE, and ROC). The analysis' findings demonstrated that all of the unique hybrid models that had been created had high predictive abilities, but the model MRAB-FT performed the best in terms of recognizing groundwater potential zones.

Tamiru et al. [15] assessed the effectiveness of artificial intelligence (AI) in the Fincha catchment, Abay, Ethiopia, using geospatial analysis and GIS platforms to prospect possible groundwater zones. In this work, a Multi-Layer Perceptron (MLP) structure was used. The analysis' findings demonstrated that the AI models agreed with 96% of the identified groundwater potential zones and ground-truthing points, while the accuracy of the GIS platform model using the AHP method was 91%. Finally, it is stated that the ANN model is a valuable tool for defining potential groundwater zones where the cost of direct field investigation is not affordable.

Consequently, the main objective of the present study is to construct a novel hybrid model for groundwater potentiality mapping (GPM) using the performance of supervised machine learning (ML)-based techniques integrated with knowledge-driven technologies using an analytical hierarchy process (AHP) to evaluate the groundwater potential zones of Dammam confined aquifer that extended along the Najaf-Muthanna governorates using supervised learning classification.

3.1 Assessment of models' predictive capability

A compression analysis is carried out for the predictive ability of the following five supervised machine learning algorithms: Ensemble Boosted Trees; Naive Bayes (NB); Support Vector Machines (SVM); Multi-Layer Perceptron (MLP), and k-Nearest Neighbours (kNN) was applied to the task of supervised classification using spatially constrained remotely sensed data. The performance of predictive models is evaluated using the metrics mentioned in section 2.5 of predictions evaluation.

3.1.1 Performance evaluation of Ensemble Boosted Trees-logic-based learners

Ensemble learning is a machine learning approach that involves training several learners to solve the same problem. Unlike traditional machine learning algorithms that attempt to learn a single hypothesis from training data, ensemble methods try to create and combine several theories. In data mining, classification algorithms can process significant information simultaneously. It is possible to establish assumptions about categorical class names, categorize information based on training sets and class labels, and classify newly obtained data using this method, mainly used for grouping purposes [30, 31]. In this case, the number of learners was equal to (30) with the learner type of decision tree. The maximum number of splits was (20) used to fit the model on the training set and evaluate it on the testing set. The accuracy of this model for training data is 72.3%. The confusion matrix plot is presented in Figure 11 to understand how the currently selected classifier is performed in each class and to identify the areas where the classifier has performed poorly.

The columns represent the predicted class, while the rows indicate the true (actual) class. The diagonal cells show where the true class and predicted class are matched. If the diagonal cells in this pattern are blue, the classifier correctly identifies the data in this true class, while classes with orange color are misclassified. On the other hand, the proportion of correctly classified observations per true class (TPR) and the proportion of incorrectly classified observations per true class (FNR) are presented in Figure 12. The last two columns on the right illustrate a percentage summary per true class. The correctly identified wells with pumping rates below 3 l/sec indicated as class 1 (very poor) are 47, as shown in Figure 11, while the misclassified classes are 4.

As a result, the TPR and FNR equal 92.2% and 7.8%, respectively, as presented in Figure 12 and calculated from Eqns. (5)-(6). In Figure 13, the result is plotted per predicted class instead of true class to investigate the False Discovery Rates FDR, the fraction of observations that were not properly categorized for each predicted class. Since the true positives (TP) observations are 47 well and false positives (FP) observations are 6 wells per predicted class. The Positive Predictive Values (PPV) and False Discovery Rates (FDR) are calculated as in Eqns. (7)-(8), which equals 88.7% and 11.3%, respectively.

The Receiver Operating Characteristic (ROC) curve represents the model's performance after training. The ROC curve displays the provided trained classifier's true positive and false positive rates. The optimal solution is a right angle that reaches the upper left corner of the plot since there are no misclassified points. A poor result may be found below a line sloped at 45 degrees, referred to as the line of a random guess. The number that represents the Area Under the Curve is a measurement of the overall quality of the classifier. The larger area under curve values indicates better classifier performance. The AUC for the Ensemble Boosted Trees model is equal to (0.98), as presented in Figure 14.

11.png

Figure 11. Confusion matrix of Ensemble Boosted Trees model

12.png

Figure 12. Ensemble Boosted Trees model performance per true class using True Positive Rates (TPR) and False Negative Rates (FNR)

13.png

Figure 13. Ensemble Boosted Trees model performance per predicted class using the Positive Predictive Values (PPV) and False Discovery Rates (FDR)

14.png

Figure 14. ROC curve of the Ensemble Boosted Trees model

15.png

Figure 15. Confusion matrix of testing data for the Ensemble Boosted Trees model

Finally, the model is tested using the 30% of remaining data that is not involved in model building and training. The accuracy of the Ensemble Boosted Trees model is calculated using Eq. (11) from the confusion matrix of the testing data as the percentage of true results in the dataset's diagonal cells (true positive and true negative) to the total number of (107) testing wells, which is equal to 89.7%, as presented in Figure 15, where R stands for the confusion matrix of the testing data and (ans) denotes the calculated accuracy value.

3.1.2 Performance evaluation of support vector machine (SVM) algorithms

By using kernel functions, the Support Vector Machine (SVM) can locate nonlinear decision boundaries [32]. There are two primary steps. In the first step, the inputted original data are initially transformed into a space with a higher dimension using the kernel function (polynomial, Gaussian, radial basis function, exponential radial function, Multi-Layer Perceptron, etc.). The second step is searching for a hyperplane that linearly separates data in the new space [33]. In the current study, medium Gaussian SVM was applied for training data; the kernel function used for data transforming into the higher dimension to generate the Gaussian surface makes the data linearly separable. The accuracy of this model for training data is 72.7%. The performance of the SVM classification learning model was evaluated using the confusion matrix and the statistical metrics: True Positive Rates (TPR), False Negative Rates (FNR), Positive Predictive Values (PPV), False Discovery Rates (FDR), and AUC are presented in Figure 16. TPR and PPV should be high, and FNR and FDR should be low for better performance. The AUC of the SVM is (0.99). On the other hand, the SVM classifier's accuracy in the case of testing data was 77.6%.

16-1.png

16-2.png

Figure 16. Performance of the SVM classification learning model: (a) confusion matrix in the form of observations number; (b) TPR and FNR ratio; (c) PPV and FDR (d) ROC curve

3.1.3 Performance evaluation of Naive Bayes (NB)-statistical learning algorithms

The methods in this category include Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA), which employ density estimation strategies to identify linear (or quadratic) combinations of variables that optimally divide groups into separate categories [34]. The Gaussian Naive Bayes classifier was used for training data in this study. It is used for continuous numerical features. The distribution of continuous values is assumed to be Gaussian/normal distribution; it gives a bell-shaped curve that is symmetric about the mean of the feature values. And therefore, the probabilities are computed based on Gaussian distribution. The performance of the Naive Bayes (NB) classifier was evaluated, as presented in Figure 17.

The accuracy of this model for training data is 66.5%. Figure 17 (b) shows that the True Positive Rate (TPR) for class 1 is 100%, indicating a perfect outcome with no misclassified points; however, the TPR for classes 3 and 5 is lower than the (FNR); therefore, the data must be trained using another classifier that provides better performance.

The AUC of the Naive Bayes classifier is equal to 0.99. In comparison, the accuracy of the current classifier for testing data is 75.7%.

17-1.png

17-2.png

Figure 17. Performance of the Naive Bayes classification learning model: (a) confusion matrix of observations number; (b) TPR and FNR ratio; (c) PPV and FDR; (d) ROC curve

3.1.4 Performance evaluation of Multi-Layer Perceptron (MLP)-Artificial Neural Networks (ANN)

Perceptron is an algorithm for supervised learning consisting of four major components: input values, weights and biases, net sum, and an activation function; Figure 18 presents these principal components, a weighted sum of inputs followed by an activation function that can be generalized to the form of:

$\mathrm{y}_{\mathrm{i}}=\mathrm{f}_{\mathrm{k}}\left(\sum_{\mathrm{i}} \mathrm{W}_{\mathrm{i}} \mathrm{X}_{\mathrm{i}}\right)$      (12)

where, W_ji represents the weight that may be adjusted for the ith instance and X_i represents one of the variables that are supplied. The activation function, denoted by f_k, may take the form of any nonlinear function, such as a step function or a sigmoidal function [23].

The trained MLP neural network of five class classification was formed by an input layer with seven neurons representing the seven data features of the thematic maps for the affecting factors of the groundwater potential map, two hidden layers with five neurons in each layer, and an output layer with five neurons. The output layer performs the required task, such as prediction and classification. The target of the network was indicated by the neurons that are output. The neurons in the MLP were trained with the backpropagation learning algorithm, and the hyperbolic tangent function is the activation function of the hidden layers. The softmax transfer function was used inside the output layer. Because of its ability to convert input data into probability, this transfer function is often used in the output layer of classified Artificial Neural Networks (ANN). The softmax normalizes an input value into values that follow a probability distribution totaling up to 1. It is preferred in the multi-class classification of the neural network model because the output values are in the range [0,1], unlike the binary classification, which can only accept the value 0 or 1. In the case of small or negative inputs, the softmax turns it into a small probability, and if the input is large, it turns it into a large probability, but it will always remain between 0 and 1 [35].

The accuracy of correct classified classes is 91.7% for (100000) epoch/iterations; the training confusion matrix of the MLP neural network is shown in Figure 19.

The rows represent the predicted neural network class (the output class), whereas the columns represent the actual class (target class). The observations that have been successfully categorized are shown by the diagonal cells. The observations that were not correctly categorized are shown by the off-diagonal cells. Each cell displays the overall number of observations and a percentage of the total observations. The percentage numbers shown in green reflect the proportion of correctly classified observations to the total observations.

In contrast, the percentages displayed in red represent the proportion of data incorrectly classified. The column on the far right of Figure 19 displays the percentages of all the predicted samples to belong to each class that was correctly and incorrectly classified. These metrics are often called the precision (or positive predictive value) for correctly classified classes and the false discovery rate for the incorrectly classified classes. The row at the bottom of the graph shows the proportion of samples properly and wrongly identified for each target class. Recall (or true positive rate) and false negative rate are common names for these statistics. The total accuracy is shown in the cell located in the plot's bottom right corner (91.7%).

Acceptable results are achieved due to the almost complete straightening ROC curve for classes 3 and 4, as presented in Figure 20.

When the area under the ROC curve is closer to the value 1 or when the ROC curve straightened at the top of the graph in the case of 100% of TPR (sensitivity) and 100% of FPR (specificity), performance levels are considered to be at their highest. Hence, the performance of classifications developed by ANN provides a perfect result for classes 1, 2 and 5. Finally, the model is tested using a 30% testing dataset to calculate the model accuracy, which was 79.4%.

18.png

Figure 18. Structure of proposed five-class MLP neural network

19.png

Figure 19. Training confusion matrix of the MLP neural network

20.png

Figure 20. Training ROC curve for MLP neural network model

3.1.5 Performance evaluation of k-Nearest Neighbours (kNN)-instance-based learners

The k-Nearest Neighbors (kNN) method is a non-parametric supervised learning technique. In this study, we investigate the impact of distance metrics while using kNN as the classifier for image classification. Non-parametric algorithms mean it makes no assumptions about the data distribution [36]. When working with KNN, the distance between two data points is determined by a similarity measure, also known as a distance function. The Euclidean distance was used in this model, generated from a generalized metric known as Minkowski distance. As p=2, Minkowski distance is specific to Euclidean distance. The majority vote of the closest sample classes identifies the sample's class [34].

The weighted kNN with the number of neighbors equal to (10) was used to train the dataset. The accuracy of the trained model is 93.0%. Figure 20 shows the current model's performance. The total number of misclassified cells is (17) knowing the total number of the trained dataset is (242).

The wells of classes 1 and 2 are perfectly classified (100%) of TPR and (0%) of FNR, while for classes 3, 4 and 5, the TPR was equal to (88.9%, 90.6%, and 89.5%) respectively. On the other hand, the kNN classifier's accuracy in the testing phase was equal to 90.7%.

The groundwater potential zones (GWPZs) in the southern desert of Iraq along the Najaf and Muthanna governorates are identified in this study, applying remote sensing (RS), geographic information system (GIS), and analytical hierarchy process (AHP) methods. With the help of published maps for soil and main groundwater aquifers units, Rs data (satellite images and STRM DEM), data from groundwater wells, and assigned weights and ranks for groundwater potentiality influencing factors and their respective classes using the AHP technique, thematic layers of aquifer unit types, transmissivity, lineaments density, slope, soil type, land use and cover, and drainage density were prepared. The prepared thematic layers were used as input variables for MLAs, and the collected data from (349) groundwater wells were used to train and evaluate the five developed hybrid ML models. The True Positive Rates (TPR), False Negative Rates (FNR), Positive Predictive Values (PPV), False Discovery Rates (FDR), the Area Under the Curve (AUC) of the Receiver Operating Characteristics (ROC), and the accuracy of the classifier are used to assess the performance of ML models. According to the analysis of ML models and based on their accuracy matrices, the results show that all developed models are highly predictive, but the k-Nearest Neighbors (kNN) model is the most effective at identifying areas with significant groundwater potential with an accuracy of 90.70%, followed by Ensemble Boosted Trees-Logic-Based Learners, Multi-Layer Perceptron (MLP)-Artificial Neural Networks (ANN), Support Vector Machine (SVM), and Naive Bayes (NB)-Statistical Learning Algorithms with validation accuracy of (89.7%, 79.4%, 77.6%, and 75.7%) respectively. In addition, the total number of misclassified cells is (17) in the kNN model, given that the training dataset size is (242). Class 1 and 2 wells are ideally rated (100% TPR and 0% FNR), while classes 3-5 had TPRs of 89%, 90%, and 89.5%.

Consequently, a five-class groundwater potential map is created using the k-Nearest Neighbors (kNN): (very poor) for pumping rates below 3 l/sec, (poor) for pumping rates between 4 and 6 l/sec, (moderate) for pumping rates between 7 and 9 l/sec, (good) for pumping rate between 10 and 14 l/sec and (very good) for pumping rates above 14 l/sec. According to the findings, there are five different groundwater potential zones (GWPZ): very poor (4281.69Km²), poor (3253.47Km²), moderate (27385.90Km²), good GWPZ (32319.82Km²), and very good (9491.84Km²).

The developed hybrid model can accurately predict the state of groundwater potential zones and allows the opportunity to characterize the behavior of a groundwater source by integrating the effect of the set of influencing factors represented by the seven thematic maps with their respective effect weights. Also, it can be modified with new data. Also, the process used in the present research could be applied and evaluated in other areas to produce GPMs. In brief, GPMs might greatly assist water resource managers and planners in understanding water resources' conditions, exploitation, and conservation measures.