The data come from ten of the fourteen regions of Senegal and The Gambia. Senegal, located in West Africa, and the Gambia at its center, as shown in Figure 1 .

These countries share the same geological heritage. Indeed, Gambia is an enclave in the center of Senegal, where geological continuity ensures the presence of the same basin. Soil samples are taken and tested in the lab to determine their chemical element content.

The samples taken from different regions depend on the diversity of colors and especially the types of materials, most often clayey, found in these different regions.

Hence, this great variability of samples depends on the region and needs. Figure 2 shows the quantities sampled by region.

Thus, depending on the soil type, we note the considerable contribution of data on clays, particularly black, to other types of soils. We show in Figure 3 the quantities of raw material in our samples according to the soil type.

This difference can be explained by the fact that the large quantity of data collected, especially in the south, is of lake origin. This environment is strongly marked not only by a dancing hydrographic network. But also, by a strong vegetation decomposition of living organisms, making these lands very rich in organic matter.

This distribution, Table 1 , broadly represents the data used to carry out the studies in this article and is distributed as follows:

In this table:

This data allows us to see the different ranges of variation by soil type and their mineral content. Sands have a high silica content and low alkaline and ferric mineral content, unlike clays. The geographical diversity of soils in Senegal highlights their mineralogical variability depending on their nature. It shows that sandy soils, which are widespread in certain regions, are richer in silica, while clay soils, which are more alkaline, contain about half as much basic oxides. This chemical distinction between soil types is essential for understanding their mechanical, physical, and biochemical behavior.

The data contains the content of silica (SiO₂), alumina (AL₂O₃), titanium monoxide (TiO), quicklime (CaO), sodium oxide (Na₂O), hematite (Fe₂O₃), magnesia (MgO), and potassium oxide (K₂O). Using this data, we calculated: the silica saturation index (IndSatSilice), soil alkalinity (indAlcalin), the proportion of iron relative to other oxides (IndFer), basic potential (SomOxyBasik), and total composition (Compotot). An initial classification is based on regions. The second classification is based on soil type, without considering the region of origin.

Senegal is a flat country whose altitude rarely exceeds 40 m ( Maignien, 1965 ) in the sedimentary basin. This flat terrain sometimes has certain consequences, such as flooding in combination with land use. This basin covers a large part of the country, unlike the Precambrian basement. It is an outcrop of hard rock in eastern Senegal, but especially in its eastern part. In this area, hills rise to a height of around 400m. The sedimentary basin provides a wealth of knowledge about the pedology of Senegal. It extends from Mauritania to Guinea-Bissau.

As part of this study, data were collected from field samples taken in several targeted regions. Sampling sites were chosen along riverbanks, areas that are particularly representative of alluvial deposits and local hydrogeological dynamics. The collection operations were carried out using a pickup truck, which allowed access to rural areas that are difficult to reach. The tools used included a GPS, a shovel, a pickaxe, an auger, and bags for packaging the materials. Each sample was collected according to a standardized protocol to ensure the representativeness and traceability of the samples. They were collected at different depths, taking into account their spatial and temporal variability.

After collection, the samples underwent a series of laboratory tests to determine their chemical composition (calcium, titanium, silica, etc.). These results were compiled in Excel, allowing for structured organization of the data. These files were then processed using Python scripts, including data cleaning for management purposes and anomaly detection. As part of this study, min-max normalization was used to control deviations, standardize variable scales, and improve algorithm convergence. Where applicable, exploratory statistical analyses reveal the influences and dominance of variables.

Support vector machines (SVMs) are supervised learning algorithms used for classification, regression, and anomaly detection problems. They are increasingly used in geoscience, particularly in geophysics, for the classification of geological formations. Determining the estimator parameters is a convex optimization problem. Any local solution to such a problem will be globally optimal ( Agrawal, Barratt, & Boyd, 2021 ). It is an estimator that uses labels. These labels are characterized by a set of features.

Each feature is a data vector used as a training set to construct a hyperplane separating the classes of a dataset. SVM is a binary classifier. For complex cases, SVM transforms the data into a higher-dimensional space to make it linearly separable. To understand the basics of SVM ( Kim et al., 2003 ), we will illustrate the approach using a binary classification problem. Suppose we have a training set ${\left\{\left(x_i,y_i\right)\right\}}_i{}_{=1,2}$ in a two-dimensional plane, where the $x_i$ are descriptors of a class $y_i$; $y_i\in \left\{-1,+1\right\}$. If the set is linearly separable, then there exists a group of linear separators called separation hyperplanes. This is valid for N-dimensions or N descriptors ( Kovacevic et al., 2009 ). Thus, these hyperplanes are modeled by the function:

${\omega }^{\text{T}}{x}_i+b=0$ where x is an input vector, w is a weight vector, and b is the threshold. Overall, we can write:

${\omega }^{\text{T}}{x}_i+b\ge 0\text{pour}{y}_i=+1$

${\omega }^{\text{T}}{x}_i+b<0\text{pour}{y}_i=-1$

The space between the separating hyperplane and the support points is called the geometric margin. The objective of SVMs is to maximize this margin, hence their name “wide margin separator.” If w₀ and b₀ are the optimal weights and bias, respectively, then the optimal hyperplane is defined by:

$w_0^{\text{T}}x+b_0=0$

The relationship between $x_i$ and $f\left(x_i\right)$ is defined by:

$f\left(x\right)=w^{\text{T}}x+b$

If we call $f\left(x\right)$ the functional separation margin, then the distance between the decision function and the support points can be written as follows:

$r=\left|\frac{f\left(x\right)}{\Vert w_0\Vert }\right|$

This relationship assumes that all training points are classified on or behind the hyperplanes. The distance between the two support vectors is defined by d:

$d=2r=\frac{2}{\Vert w\Vert }$

The objective function can be represented by:

$\max \frac{2}{\Vert w\Vert }{y}_i\left(w^{\text{T}}{x}_i+b\right)\gg 1,i=1,\cdots ,n$

The solution to such an optimization problem is given by the Lagrange function.

$L\left(w,b,\alpha \right)=\frac{1}{2}{\Vert w\Vert }^2-{\displaystyle \sum {\alpha }_i\left[y_i\left(w^{\text{T}}{x}_i+b\right)\right]}$

where ${\alpha }_i$ are Lagrange multipliers. This function is minimized by the values of w and b and maximized by the positivity of the multiplier coefficients. The multiplier coefficients are determined by the following optimization function:

${\alpha }_i-\frac{1}{2}{\alpha }_i{\alpha }_j{y}_i{y}_j\left(x_i{x}_j\right)\text{avec}{\alpha }_i\gg 0,i=1,\cdots ,l\text{et}{\alpha }_i{y}_i=0$

It relies on a kernel to determine a separating hyperplane. This hyperplane can be linear if no mixing of data is observed. It can be flexible if a certain degree of tortuosity is observed in the separating hyperplane. It will be associated with another data separation model called “One versus All”, of which another variable, “One versus One”, exists.

The “one versus all” approach results in one classifier per class, for a total of five classifiers. For class i, it will consider the labels of that class as positive and the others as negative.

Whereas “one versus one” forms a separate classifier for each pair of different labels. This leads to 5(5-1) / 2 = 10, or 10 classifiers. However, despite the high number of models, this method is much less sensitive to data set imbalance issues and is much more computationally expensive.

In soil studies, data cannot generally be separated in a linear fashion, as decision boundaries are often complex and non-trivial. Indeed, the characteristic properties of a soil type may appear locally in the form of inclusions, lenses or dome-shaped structures within the same geological layer, making classification more difficult. This spatial heterogeneity requires the use of models capable of capturing the non-linear relationships between variables.

Thus, these learning data are projected into an N-dimensional space corresponding to the size of the descriptors ( Liu & Xu, 2013 ). Depending on the flexibility of the kernel and the hardware and software resources chosen, they can be separated as best as possible. The following kernels are used:

$K\left(x_i,x_j\right)=x_i^{\text{T}}{x}_j$

$K\left(x_i,x_j\right)={\left(\gamma x_i^{\text{T}}{x}_j+r\right)}^d,\gamma >0$

$K\left(x_i,x_j\right)=\text{exp}\left(-\gamma {\Vert x_i-x_j\Vert }^2\right),\gamma >0$

$K\left(x_i,x_j\right)=\text{tanh}\left(\gamma x_i^{\text{T}}{x}_j+r\right)$

The values γ, r, and d are parameters of the kernel in question. In geophysics, the RBF kernel is the most commonly used ( Abedi, Norouzi, & Bahroudi, 2011 ):

SVMs have proven successful in classification problems. This is largely due to parameter selection ( Agrawal, Barratt, & Boyd, 2021 ). For a linear model, defining the regularization parameter C is sufficient for good classification. Unfortunately, data are not always linearly separable, particularly in structural geology, where layers are not horizontal. The parameters d for a polynomial kernel or γ for the RBF are added to the regularization parameter. In order to find the most optimal values for these parameters, the grid search method was introduced. It allows us to find the best combination of parameters for our problem.

The data was divided according to the following percentages in both cases: training: 80%, testing: 15%, and 5% set aside to verify the model ( Agrawal, Barratt, & Boyd, 2021 ; Kovacevic et al., 2009 ). In the first case, raw values are used, and in the second, the data is normalized. To improve the performance of our different models, we varied the search criteria for training. The following performance metrics: overall precision or accuracy, true positive rate, recall, and F-1 score are used to evaluate the performance of our models ( Ruano et al., 2013 ).

This study is based on a dataset limited to specific geographical regions and well-targeted sectors within those regions. This may restrict the scope and generalization of the results to other geological contexts. Indeed, the physical and chemical characteristics of soils vary considerably depending on local climatic, geological and anthropogenic conditions. For example, a model trained on lateritic clays and black clays in tropical areas may not be directly applicable to marl-limestone or sandy soils in temperate environments, due to marked structural and behavioral differences.