The study site is located in the Lower Tassaout region of Morocco ( Figure 1 ), an irrigated agricultural area where soil salinity has been a long-standing challenge, exacerbated by irrigation practices over the past five decades. The region experiences a semi-arid climate characterized by low and erratic rainfall, high evapotranspiration, and frequent droughts.

According to the World Reference Base (WRB) classification (19), the dominant soil types include Solonetz, Luvisoils with Solonetz tendency, and Solonchaks. The area is predominantly used for irrigated farming, producing cereals (e.g., wheat and barley), vegetables (e.g., tomatoes, onions, and peppers), and fruit crops (e.g., citrus fruits, olives, and melons). Forage crops like alfalfa are also widely cultivated to support livestock. Water for irrigation is primarily sourced from groundwater, with supplemental inputs from seasonal floods and small-scale reservoirs or ponds constructed for water harvesting (20).

A representative section of the Lower Tassaout irrigated perimeter was selected as the study area based on preliminary field surveys, historical land use assessments, and consultations with local farmers and agricultural extension agents. The selected zone was chosen for its diversity in soil types and salinity levels, making it suitable for evaluating spatial variability in salinity conditions.

To effectively capture this variability, a 100 m × 100 m grid was superimposed over a 1 km² area, resulting in 103 sampling locations (S01–S111). A stratified random sampling approach was employed, considering both soil type distribution and known salinity gradients to ensure balanced spatial coverage and representative sampling across soil and salinity classes.

At each grid cell, a composite surface soil sample (0 – 20 cm depth) was collected for physico-chemical analysis, including EC_S:W and EC_e. Additionally, seven soil profile pits (P1–P7) were excavated across different zones to assess vertical salinity variation and support the calibration and validation of predictive models. This integrated design provided a comprehensive dataset to evaluate the potential of pXRF spectroscopy as a rapid, non-destructive tool for predicting EC_e in salt-affected soils.

The soil samples were oven-dried at 40 °C to remove excess moisture as described in the ISO 11464 (21) protocol. The dried soil is then ground using a mortar and pestle and sieved through 2-mm mesh to exclude large particles and debris and to maintain the uniformity of samples. For the saturated paste extracts used for EC_e measurement, 200 g of soil was weighed into a 500 ml ceramic dish and mixed with deionized water until a saturated paste was formed, following the method described by Rhoades (22). The paste was carefully observed for the development of a glistening characteristic when light was reflected on it and checked for flowability when the container was tipped. The pastes were allowed to stand overnight until they reached equilibrium. Once prepared, the pastes were placed in mechanical extractors with filter papers on top and syringes attached to connection tubes. The extractor was set to run for one hour to extract the soil paste extracts. Finally, EC_e was measured using a calibrated EC meter (SevenDirect SD30 – Mettler Toledo). In accordance with ISO 11265 (1999) protocol, three soil-to-water suspensions were prepared by weighing 10 grams of soil in 50 ml polypropylene tubes and adding 10, 25, and 50 ml of water, respectively. The tubes were then tightly sealed and shaken at 20 °C for 30 minutes at a speed of 180 rpm while ensuring the soil samples are mixed well with water. Subsequently, the tubes were immediately centrifuged at 3500 rpm for three minutes until supernatant liquid was separated from the soil particles. Finally, EC_S:W measurements were performed in duplicate using the same EC meter used for EC_e determination.

Soil samples for analysis using the pXRF technique in the laboratory were prepared by grinding and sieving through a 180-µm nylon mesh sieve, while the pXRF technique itself remains non-destructive to the sample. The device used for soil measurement was a Tracer 5I (Bruker) portable X-ray fluorescence instrument, which was set to operate at a voltage of 10 keV and a current of 70 µA to excite the elements of interest. Besides, an 8 mm collimator used for providing a focused beam for accurate reading, and to minimize background radiation interference and enhance target element recognition, a blank manual filter was employed. Furthermore, the calibration of the instrument was verified using a conventional metal coupon device. In preparation for measurement, a Soil Light option was chosen in the menu of the device because the Soil Light option is specifically designed for measuring light elements such as calcium (Ca), sodium (Na), magnesium (Mg), and chlorine (Cl), which are key elements for soil salinity and sodicity characterization. Subsequently, measurement of elements was conducted. Each measurement was conducted over a 90-second period to ensure sufficient counting statistics and reliable results are achieved. Figure 2 depicts the workflow of soil measurement using a pXRF technique.

To evaluate the potential of pXRF spectroscopy for predicting soil salinity, especially EC_S:W and EC_e, two multivariate statistical approaches were employed: principal component analysis (PCA) (23) and partial least squares regression (PLSR) (14). These were applied to explore patterns in the spectral data and to develop predictive models based on elemental signatures related to salinity (24, 25).

PCA was conducted using the “stats” package in R version 4.5.0 (26) and served as an exploratory step to reduce the dimensionality of the pXRF dataset while retaining most of the spectral variability. It transforms correlated spectral variables into uncorrelated principal components (PCs), enabling the identification of sample groupings, outliers, and spectral variance patterns (27). These components also help interpret underlying geochemical gradients related to salinity.

Following PCA, PLSR was used to model the quantitative relationship between the full pXRF spectra (predictor matrix) and the measured EC values (response variables). PLSR is especially suited to spectral data due to its ability to handle high-dimensional and collinear predictors. It projects both predictors and responses onto a new set of latent variables that maximize their covariance, effectively capturing the underlying structure in complex datasets (25).

To avoid overfitting and optimize model performance, the number of latent variables was selected using k-fold cross-validation, with model selection guided by the minimization of the Root Mean Square Error of Cross-Validation (RMSE_CV) and the maximization of R². Once optimized, final models were trained on 75% of the dataset and evaluated on a 25% independent validation set.

Model performance was assessed using three key metrics:

where y_i are the observed values, y ^ i are the predicted values, y ¯ i are the mean of the observed values, n is the number of observations, IQR(y) is the interquartile range 25% - 75% of the observed values. Models with higher R² and RPIQ values and lower RMSE were considered superior in predictive accuracy and robustness.

Additionally, other statistical tests were performed to assess the assumptions and robustness of the models. The Shapiro-Wilk test was conducted to evaluate the normality of residuals, ensuring that the assumptions of linear regression were met. Levene’s Test was used to check for the homogeneity of residuals, confirming that the variance of residuals was consistent across different groups. The Kruskal-Wallis test was applied to test for significant differences between multiple EC groups, particularly when assumptions of normality were violated. Finally, Dunn’s test was used for post-hoc comparisons between EC groups to identify specific differences between them. Together, this modeling and validation framework ensured a comprehensive assessment of predictive performance, addressing not only predictive accuracy but also model robustness, residual behavior, and statistical integrity.

Table 1 shows the summary statistics for the soil EC_S:W. For the 1:1 soil-to-water ratio, the electrical conductivities ranged from 0.33 to 77.88 dS.m^-1, while for 1:2.5 and 1:5 ratios, values ranged from 0.17 to 38.87 dS.m^-1 and from 0.09 to 20.96 dS.m^-1, respectively. As expected, EC values decreased progressively with higher dilution ratios due to the reduction in ion concentration. The average EC dropped from 10.85 dS.m^-1 at 1:1 to 2.67 dS.m^-1 at 1:5, confirming this trend.

The equations generated from the linear regressions, shown in Figure 3 , demonstrate strong linear correlations between the various dilution ratios with high R² values (all above 0.99). This trend, based on EC values, aligns with findings from Klaustermeier et al. (28), who reported an R² value of 0.91 for the relation between EC_1:1 and EC_1:5. The regression equations indicate a predicted decrease in EC with higher dilution ratios, as evidenced by decreasing slopes corresponding to increasing dilution. The minimal intercepts, approaching zero, suggest that the linear models possess negligible bias, enhancing the reliability of EC predictions across varying ratios. This capability facilitates the estimation of EC at multiple soil-to-water ratios from a single measurement, which is particularly advantageous for rapid or approximate salinity assessments where time or resources are limited. Overall, these findings demonstrate the effectiveness of linear regression models in accurately capturing the relationship between EC values across different dilution ratios.

The plots in this study ( Figure 4 ) reveal a strong linear relationship between EC_S:W (1:1, 1:2.5, and 1:5) and EC_e, with high R² values of 0.963, 0.944, and 0.950, respectively, confirming that EC_S:W is a reliable predictor of EC_e. As the dilution ratio increases, the regression slope becomes steeper, indicating that EC readings decrease more relative to EC_e, thus requiring larger scaling factors for accurate conversion. This trend reflects the dilution effect, where increased water reduces ion concentration, amplifying the difference between EC_S:W and EC_e. The presence of positive intercepts in each regression suggests a baseline conductivity in the saturated paste that is not influenced by dilution. These results highlight the importance of selecting the appropriate regression equation for each ratio; while the 1:1 ratio shows a more direct relationship, higher ratios like 1:5 require greater adjustment for accurate EC_e estimation.

Table 2 presents findings from other studies, and our regression models show alignment with these previous results, though with some variations in slopes and interceptions due to methodological differences. For the 1:1 ratio, our regression model EC_e=2.13*EC_1:1 + 0.76 with R² = 0.96 is comparable to previous studies, such as Sonmez et al. (33) with a similar slope of 2.03 and an intercept of -0.41, as well as Zhang et al. (31), who reported EC_e=1.79*EC_1:1 + 1.46. However, our slope is higher than those reported by (30) and (34), which found slopes of 1.56 and 1.83, respectively, indicating stronger EC_e sensitivity in our dataset. This variability could stem from differences in extraction methods and soil compositions across studies.

For the 1:2.5 ratio, our model EC_e=4.33*EC_1:2.5 + 2.3 with R²=0.94 aligns well with (39) for similar ratios, who reported slopes of 3.719 and 3.514 for the 1:2 ratio with R² values of 0.85 and 0.91. The slightly higher slope in our model indicates an increased sensitivity of EC_e to EC_S:W in this ratio, possibly due to the Rhoades method used, which is known to produce precise results for soil salinity.

For the 1:5 ratio, our regression EC_e=8.14*EC_1:5 + 2.2 with R²=0.95 is within the range of slopes reported by previous studies. For example, Sonmez et al. (33) reported a slope of 7.36 with an exceptionally high R² of 0.99, while other studies such as Ozcan et al. (46) and Aboukila et al. (44) reported slopes around 5.97 and 7.46, respectively. Our slope is relatively high compared to these, indicating that EC_e is more sensitive to changes in EC for the 1:5 ratio in our dataset. This increased sensitivity aligns with findings from Marien et al. (45), who observed a slope of 11.35, emphasizing the potential for variability due to differences in soil type, extraction methods, and EC range.

Overall, our study’s results show that EC_e can be accurately estimated from EC_S:W, with differences in slopes and intercepts reflecting methodological diversity across studies. These findings reinforce the importance of using context-specific calibration equations for accurate soil salinity assessments.

Figure 5 presents the pXRF spectra of representative soil samples, showcasing the elemental composition through distinct fluorescence peaks. Key elements detected include magnesium (Mg), aluminum (Al), silicon (Si), sulfur (S), chlorine (Cl), potassium (K), calcium (Ca), manganese (Mn), and iron (Fe), corresponding to their characteristic Kα₁ X-ray energies at approximately 1.48 keV, 1.74 keV, 2.31 keV, 2.62 keV, 3.31 keV, 3.69 keV, 5.90 keV, and 6.40 keV, respectively. Among these, Ca exhibits the highest peak intensity at around 4 keV, indicating its substantial presence in the samples. This strong signal is consistent with the role of Ca in maintaining soil structure and influencing salinity, as calcium salts such as CaCl₂ and CaSO₄ are commonly found in arid and semi-arid soils. The prevalence of Ca in these soils also reflects its involvement in cation exchange processes and aggregation stability, both of which are crucial under saline conditions.

Although sodium (Na) is a key driver of soil salinization and sodicity, it was not detected in the pXRF spectra. This is due to the instrument physical limitation: Na emits very low-energy X-rays (~1.04 keV) that are strongly absorbed by air and the detector window, making it difficult for handheld pXRF devices to quantify Na reliably. Consequently, Na was excluded from the measurements. This limitation is well documented in pXRF applications (47), and highlights the need to complement pXRF analysis with conventional laboratory methods when Na quantification is required.

Fe peak near 6.4 keV is also prominent, suggesting a considerable iron content across the samples. While Fe is not directly linked to salinity, its presence influences clay mineralogy and soil fertility, potentially affecting the soil capacity to retain or mobilize salinity-related ions. This is particularly relevant in soils where Fe oxides or Fe-rich clays interact with other salts or affect redox-sensitive elements.

Of particular interest are the peaks of Cl at ~2.6 keV and S at ~2.3 keV, both of which are well-recognized markers of salinity-inducing ions. The high Cl signal suggests accumulation of halite (NaCl) or other chloride-based salts, which are known to significantly elevate soil EC (48). Similarly, the presence of S, often linked with sulfate salts such as gypsum (CaSO₄·2H₂O), further reinforces the connection between the pXRF spectra and salt composition. In agreement with prior studies, both Cl and S show positive correlations with EC (24), reflecting the enrichment of Cl⁻ and SO₄²⁻ ions in the soil and their role in determining salinity status.

The lower intensities of Si and Al further suggest the presence of aluminosilicate minerals, such as quartz and feldspars, which are common in soil matrices and contribute to the structural framework. However, these elements play a minimal direct role in determining salinity when compared to more reactive ions like Ca, Cl, and S.

Altogether, the spectra illustrate the complex interplay between elemental composition and soil salinity. While Cl and S peaks directly inform EC estimation, elements like Fe and Ca reflect both salinity influence and mineralogical background, with intensity variations pointing to differences in parent material or salt accumulation across samples. These findings underscore the utility of pXRF as a rapid, non-destructive tool for simultaneously detecting multiple key elements, enabling a better understanding and management of soil salinity across diverse agricultural and environmental settings.

To estimate soil EC_S:W under various extraction ratios and EC_e from pXRF spectra, Partial Least Squares Regression (PLSR) was selected as the modeling approach due to its demonstrated ability to handle high-dimensional and collinear datasets, characteristics typical of spectral data like that produced by pXRF. In addition to its suitability for such data structures, PLSR is widely supported in the literature as a reliable method for soil salinity prediction. For instance, Naimi et al. (49) evaluated PLSR models for EC_S:W prediction in arid Iranian soils using pXRF data, reporting coefficients of determination (R²) ranging from 0.62 to 0.84, depending on the EC measurement method and dataset partitioning. Similarly, Gozukara et al. (14) assessed multiple machines learning algorithms, including Support Vector Regression (SVR), Random Forest (RF), Cubist, k-Nearest Neighbors (KNN), and PLSR on Turkish soils and found that PLSR achieved R² values between 0.34 and 0.86, ranking among the top-performing models. These findings reinforce the strength of PLSR, which, despite its linear structure, has consistently shown competitive predictive accuracy across diverse salinity contexts. While the current study did not include an internal comparison with alternative machine learning algorithms, the strong and repeatable performance of PLSR in previous studies supports its use here as a robust, interpretable, and well-established baseline for estimating EC_S:W and EC_e from pXRF spectra ( Table 3 ).

Before applying predictive models, principal component analysis (PCA) was conducted to explore the underlying structure and variance within the pXRF spectral dataset. The PCA proportion of variance ( Figure 6 , left panel) shows that the first two principal components (PC1 and PC2) explain 30.3% and 9.4% of the total variance, respectively, together accounting for nearly 40% of the dataset variability. The sharp decline in explained variance beyond PC2 suggests that the remaining components contribute minimally and are less informative for interpretation. The individual PCA plot ( Figure 6 , right panel) projects samples onto the PC1 and PC2 plane, where most individuals cluster tightly around the origin, indicating relatively homogeneous variation. However, several samples, such as 1, 2, 3, 5, 12, and 134, appear as outliers, positioned away from the central cluster. These outliers may reflect specific conditions or anomalies, potentially related to variations in soil salinity or differences in EC measurements. The PCA effectively highlights such patterns, supporting the evaluation of pXRF-derived EC estimations (e.g., EC_S:W and EC_e) and revealing meaningful groupings within the data. Overall, the PCA results underscore the utility of this approach in reducing data dimensionality, identifying structure, and guiding further investigation into salinity-related variability among soil samples.

To gain insight into the spectral features most influential in predicting EC_S:W and EC_e, the loading plot of the first two components (PC1 and PC2) from the PLSR model was examined ( Figure 7 ). These loadings revealed distinct peaks aligned with known elemental emission lines, identifying energy regions most relevant to EC prediction. Positive loadings indicate that higher fluorescence intensity at a given energy is associated with increased EC values, while negative loadings suggest an inverse relationship.

One of the most prominent features in PC1 was a strong negative loading around 3.7 keV, corresponding to the Kα emission line of calcium (Ca Kα₁ ≈ 3.69 keV) and its associated Kβ line (~4.01 keV). This indicates that higher Ca fluorescence intensities are associated with lower predicted EC values. While calcium salts such as CaCl₂ and CaSO₄ contribute to soil salinity, the inverse relationship may reflect the stabilizing effect of Ca²⁺ on soil structure or its association with non-soluble forms. High calcium content may also reduce the mobility of other salts through buffering or precipitation processes, explaining its negative influence in the model. This region also includes a smaller negative feature at ~3.3 keV, corresponding to potassium (K Kα₁ ≈ 3.31 keV), which partially overlaps with Ca Kβ. Potassium appears to follow a similar trend, though its overall contribution is smaller. Present in both exchangeable forms and clay minerals, K may play a secondary role in EC prediction in this dataset.

In contrast, a strong positive peak was observed in the 6.4 – 7.0 keV range, corresponding to the Kα and Kβ emission lines of iron (Fe Kα₁ ≈ 6.40 keV, Kβ ≈ 7.06 keV). This suggests that higher Fe fluorescence is associated with higher predicted EC values. Although Fe is not a conventional salt-forming ion in soil solution, its correlation with EC may result from its prevalence in clay-rich or alluvial soils that retain more soluble salts. Additionally, in reducing or waterlogged environments, Fe²⁺ can become more mobile, potentially influencing salinity indirectly. Notably, Ca and Fe show opposing signs in the PC1 loadings, suggesting that they capture contrasting geochemical regimes within the dataset.

A moderate positive feature was also observed near 4.5 keV, likely corresponding to titanium (Ti Kα₁ ≈ 4.51 keV). Like Fe, Ti is commonly associated with clay minerals and may reflect underlying mineralogical differences related to salinity. The In summary, the signs and magnitudes of these elemental loadings, particularly the strong negative loading for Ca and the strong positive loading for Fe, highlight their pivotal roles in shaping EC predictions. The model’s high performance (R² ≈ 0.94 – 0.98) supports the interpretation that these elemental fingerprints, especially Ca²⁺ and Fe²⁺/Fe³⁺, are key drivers of salinity- related variability captured by the pXRF-PLSR approach.