The data comes mainly from Senegal, The Gambia and Mali. These neighboring countries share a common geological heritage. This closely related geological heritage is linked to the geological continuity of a single basin. Table 1 shows the distribution of acquisitions by country. These acquisitions are not made uniformly, but rather at sites carefully selected for their mineralogical potential.

The data collected comes from different countries but pertains to a single continuous geological zone. This zone consists of clay, sand, and granite, as shown in Table 2, which presents the quantities collected by soil type.

Table 1. Sample sizes by country.

Table 2. Sample sizes by soil type.

The samples collected reflect the diversity of soils, colors and, above all, the types of materials found in these different regions. This variation is explained by the fact that a large proportion of the data collected, particularly in the south, comes from lacustrine environments. These environments are characterized by a meandering river network, but also by significant plant decomposition of living organisms, making these soils very rich in organic matter. These data allow us to observe the ranges of variation in composition according to soil type and their mineral content. The diversity of the soils highlights their mineralogical variability depending on the conditions of formation. The data include, amongst other things: the content of silica (SiO₂), alumina (Al₂O₃), titanium monoxide (TiO₂), quicklime (CaO), sodium oxide (Na₂O), hematite (Fe₂O₃), magnesia (MgO) and potassium oxide (K₂O). Using these data, we calculated: the silica saturation index (IndSatSilice), soil alkalinity (indAlcalin), the proportion of iron relative to other oxides (IndFer), the basic potential (SomOxyBasik) and the total composition (Compotot).

In our study, the response variables correspond to the five soil types defined based on their mineralogical characteristics. The indices used provide additional information that facilitates a better understanding of the soils and their behavior.

This classification is based on soil type characteristics, without taking the region of origin into account. The Precambrian basement is an outcrop of hard rock located in eastern Senegal. This sedimentary basin provides a wealth of information on the soil science of Senegal. It stretches from Mauritania to Guinea-Bissau.

As part of this study, data were collected from several targeted regions in Senegal, The Gambia and Mali. Sampling sites were selected along riverbanks, in areas particularly representative of alluvial deposits and local hydrogeological dynamics. The collection operations were carried out using a pick-up truck. The tools used included a GPS, a spade, a pickaxe, an auger and bags for packaging the collected materials. Sampling depths varied depending on the material but generally ranged from 30 cm to 1.5 m. Sampling was carried out according to spatial variability and soil friability. As part of this study, scalar robust normalization was used to control outliers, standardize the scales of the variables and improve the convergence of the algorithms. Exploratory statistical analyses enable the data to be visualized in order to understand the influences and dominance of the variables. These collected samples are predominantly rich in clay, granite and sand. These are the major soil types present in the area. Table 3 shows the overall statistics for the data collected from the sites. It is on the basis of this data that the scaling will be applied.

Table 3. Mean and variance of the measured chemical parameters.

The compressive strength of samples taken from different soils ranges from very dense/rigid soil to very soft rock, with typical compressive strength values ranging from 0.95 to 5 MPa. Soil analysis plays a crucial role in nutrient management. Samples taken are subjected to laboratory analysis to determine their chemical composition. Chemical soil analysis serves to quantify the nutrients that should be available to plants and for human needs. Soil analysis involving the characterization of its constituents and the composition of its inorganic phases is referred to as chemical analysis. A chemical analysis involves the identification of several parameters such as: pH, redox potential, organic matter content, total nitrogen, calcium carbonate content, available phosphorus, boron, metals (Cu, Zn, Mn and Fe), exchangeable potassium, calcium, magnesium, sodium, and anion content ( NO 3 − , SO 4 2 − , PO 4 3 − , Cl⁻) ([16]). The above parameters are generally measured by soil analysis laboratories. They determine the soil category based on its mineralogical composition. The content of chemical elements is determined by inductively coupled plasma atomic emission spectroscopy (ICP-AES), with soil samples first being treated with acids. Table 1 shows the overall statistics for the data collected from the sites. It is on the basis of this data that the scaling will be applied.

Random forests are based on the principle known as the wisdom of crowds. They combine multiple decision trees built on samples obtained via bootstrapping and subsets of features ([3]). The idea is to produce robust predictions by aggregating the results of the subsets known as trees. For classification, this involves a majority vote, and for regression, an average. This approach is robust to outliers in various fields such as geophysics, geology and geochemistry. Visual geological mapping is slow and biased; random forests enable the production of more reliable maps useful for targeting areas of interest ([4]). Choosing the number of trees to ensure a balance between interpretability and algorithmic complexity remains one of the major challenges. Splitting the data into trees aims to reduce node impurity by using measures such as the Gini index or entropy. The optimal split is the one that maximizes the information gain.

Logistic regression is a primarily binary classification algorithm. It transforms a linear combination of variables into a probability using the sigmoid function. This probability is used to determine whether an example belongs to a given class. The characteristic parameters of a given sample are used to predict categorical outcomes ([5]). For multi-class problems, it is extended using the one-versus-all strategy. This strategy generates a model for each class using the softmax function ([2]). These extensions allow a probability vector, which sums to 1, to be assigned to several classes simultaneously whilst maintaining the model’s interpretability. However, performance depends on the choice of strategy and the quality of the data, particularly for imbalanced classes.

The Naive Bayes algorithm is a simple and effective classifier, well-suited to high-dimensional data such as that encountered in soil classification, particularly for food crops ([7]). Its Gaussian version (GNB) assumes that the features are continuous and follow a normal distribution. This simplifies calculations but limits its applicability to real-world data, which do not adhere to this assumption of a normal distribution. To overcome this limitation, a variant known as Stable Naive Bayes (SNB) replaces the Gaussian distribution with stable distributions, capable of modelling asymmetry and outliers in real-world data. These stable distributions generalize the normal distribution whilst retaining its property of stability under addition. This makes them suitable for real-world or non-symmetric data. This version is defined by four main parameters: α, which controls the tails or outliers; β, the asymmetry of the data; γ, their scale; and δ, the location. This offers increased flexibility for feature analysis ([19]). This extension of Naïve Bayes makes it more robust when dealing with complex distributions. Thus, SNB retains Bayesian simplicity whilst improving accuracy on real-world data.

This is an ensemble algorithm that combines weak classifiers to produce a more powerful predictive model. It operates through successive iterations, with each model being trained on weighted data. It demonstrates greater accuracy in lithology prediction ([15]). The subsequent iteration focuses on the misclassified samples from the previous one. The key hyperparameters of this estimator include the choice of base classifier, the number of iterations, and the learning rate, which controls the influence of each model. It is primarily used for binary or multi-class classification tasks. Its strength lies in its simplicity and efficiency, although it is sensitive to outliers in the measured data.

Gradient Boosting is an ensemble method based on the principle of sequentially building weak models, known as bootstrapping. This process aims to correct the prediction errors of previous training runs. These models have improved the ability to predict the risk of landslides ([18]). There is a version for regression, but in the case of classification, the algorithm used is the Gradient Boosting Classifier. Thus, compared to other algorithms, Gradient Boosting has demonstrated good performance based on wave velocities, density and natural gamma to target mineral deposits ([1]). Where each new decision tree is trained to improve the separation between classes by following the gradient of the loss function. Both models apply the same logic, but regression tasks aim to progressively reduce the residual prediction error. The final model, in both classification and regression cases, is a weighted combination of several weak trees. This allows complex and non-linear relationships in real-world data to be captured. The key parameters are the loss function, the number of trees and the learning rate. These parameters must be carefully tuned to avoid overfitting. GBC and GBR offer great flexibility and remarkable performance.

XGB models are optimized variants of Gradient Boosting, designed to be faster and more efficient in terms of computational cost and algorithmic complexity on large datasets. As with the ensemble methods mentioned, they rely on the sequential construction of weak decision trees. Each new tree corrects the errors of the previous ones by following the gradient of the loss function. XGBoost introduces regularization mechanisms (L1 and L2) to reduce the risk of overfitting whilst improving the model’s generalization. In classification, it assigns probabilities to classes. Based on class probabilities, it makes a final decision, whereas for regression, the decision is based on an optimization of numerical predictions. For estimating above-ground biomass, XGBR has demonstrated its potential in terms of accuracy compared to other methods ([13]). Key parameters include those for gradient descent as well as metaheuristic optimization. This study demonstrates the ability of XGBC to predict improvements in academic achievement through time management ([9]). These results demonstrate that ensemble models are renowned for their performance, flexibility and ability to handle complex and large-scale data. In summary, XGBC and XGBR offer a combination of power, speed and robustness, which explains their widespread adoption in research, particularly in geoscience.

The Voting Classifier is a technique based on the principle of the wisdom of crowds as applied to supervised models. It combines algorithms to produce predictions that are more robust than those obtained by a single model trained on the entire dataset. The decision is then made by majority vote (hard voting) or by combining probabilities (soft voting). Key parameters include the type of estimators and the weights assigned to them. Weighted voting is based on differential evolution, which improves classification accuracy and offers strong generalization ability as well as broad applicability ([20]). Bagging is based on the principle of stabilizing models by combining multiple predictors trained on different batches of samples drawn from the same dataset. Each subsample is obtained through bootstrap sampling, which introduces diversity and robustness into the constructed models. Bagging is a technique designed simply to reduce the prediction error of machine learning algorithms by reducing the variance of unstable prediction methods ([11]).

Before applying the machine learning model, the features must be rescaled through normalization. This involves rescaling the features in the dataset so that they are centered around the median using interquartile rescaling. Robust scaling algorithms enable features to be scaled in a way that is resilient to outliers in the measured data. This method is very similar to the MinMax scaling method, but its distinctive feature is that it uses interquartile ranges rather than the minimum and maximum values used in MinMax scaling. This is a scaling algorithm that reduces the influence of outliers by using the median and the extremes of the data based on the quantile range. For this scaling, each value of x is calculated relative to the first quartile Q₁ and then normalized relative to the interquartile range, as shown in the following equation:

An examination of the factors to interpret the underlying properties they represent shows, as illustrated in Figure 2, the following contributions by groups of variables:

Figure 2. Factors interpretation.

The variable components Fe₂O₃, MgO, CaO and IndFer reflect a dimension linked to coloring oxides and ferric indices. Soils with high levels of these components are often associated with the clays typical of central Senegal.The Al₂O₃, K₂O and Na₂O components indicate basic soils. These soils are the opposite of those dominated by compounds with a high iron index, reflecting a chemical contrast between ferric environments and the basic environments of the south.

These factors indicate ferric, basic, alkaline or siliceous soil types, distinguishing between soils rich in iron oxides and those that are more basic, magnesium-rich or rich in sodium and potassium oxides. In practice, this reveals an important geochemical dimension for soil classification. In this study, exploratory factor analysis is used to reduce the number of sample characteristics and identify the underlying factors.

Factor analysis of the data reduced the initial set of nine chemical characteristics to a set of five. These characteristics, grouped into adjacent sub-groups, enable four new characteristics to be derived for core samples based on their shared geochemical traits. These groups represent similar geochemical facies. Cluster analysis is a suitable approach for assigning a common facies label to similar samples. This clustering makes it easy to distinguish between samples. This batch analysis is a form of classification that falls broadly under the category of unsupervised machine learning. It is an approach used to infer a structure from the data measured within the dataset itself. Figure 3 illustrates the extent to which the silica and alumina content in soils in Senegal significantly influences their characterization. The data are reduced to principal components using the K-means approach, as shown in Figure 4. Each sample is thus assigned to a cluster representing geochemical facies. Examination of the signatures reveals characteristics specific to each cluster; for example, the iron index in the case of red clay soils. All these various transformations were applied to the raw data before it was divided into training, test, and validation sets.

Figure 3. Retention threshold of the eigenvalue distribution.

Figure 4. Comparison of k-means++ and random initialization.

In the context of the experiment, classification accuracy depends on the estimators’ hyperparameters and their tuning. Table 5 compares the classification results, summarizing the performance of our models in terms of precision, recall, F1-score and accuracy. We observe that the random forest achieves the best overall classification performance with a precision of 99% and an accuracy of the same value. It is followed by X gradient boosting and the bagging classifier, which are very close behind (98% overall precision each). Logistic regression and the gradient boosting classifier follow closely in terms of overall precision. The Adaboost estimator provides the lowest overall accuracy in our classification.

Table 5. Comparative analysis of supervised methods.

k = kernel; Acc = Accuracy; P = Precision; R = Recall; F1 = F1-score.