The study area covers three regions of Morocco with diverse climatic and geological conditions (Figure 1). Kenitra, located in northwest Morocco along the Atlantic coastline, experiences a maritime subhumid climate with average annual temperatures of 12.8-22.4 °C and yearly rainfall of 718.6 mm (29). The region sits on the Gharb Plain, geologically composed of Paleozoic sedimentary and metamorphic rocks overlain by Triassic red clays and Jurassic-Cretaceous dolomites (30). Fertile clay and sandy soils support cereal, olive, and sugar beet cultivation. Beni Amir, situated in the central Tadla Plain, is characterized by a semi-arid to warm Mediterranean climate with average temperatures around 20 °C and annual rainfall of approximately 430 mm (31). Its geology comprises heterogeneous Mio-Pliocene and Quaternary deposits within a synclinal depression. The predominantly clayey and sandy soils support intensive irrigated agriculture including oranges, olives, and sugarcane through the Oum Er Rbia River (32). El Jadida, part of the Moroccan Meseta, experiences a Mediterranean climate with Atlantic oceanic influence, receiving about 551 mm annual rainfall and temperatures ranging from 12 °C (winter) to 23 °C (summer) (33). The geology features tabular sedimentary layers from Tertiary, Secondary, and Quaternary periods. Sandy and clayey soils support citrus and vegetable farming. A total of 67 surface soil samples were collected at 0–20 cm depth: 21 from El Jadida, 23 from Beni Amir, and 23 from Kenitra. Composite samples (~1 kg each) were obtained by mixing subsamples using a screw auger, stored in labeled airtight bags, and transported to the Center of Excellence for Soil and Fertilizer Research in Africa (CESFRA) laboratory.

Soil samples were dried, at 40°for 48 hours in the drying oven, crushed and sieved through 2-mm sieve. Then samples were further ground using a Retsch Mortar Grinder RM 200 until a particle size of finer than 75µm was obtained. concentrations. Then the total concentrations of nine total trace elements (i.e., As, Cd, Cu, Mn, Ba, Pb, Sr, Ti, Zn) were determined by inductively coupled plasma optical emission spectroscopy (ICP-OES, Agilent 5800) following acid digestion using ultrapure HNO₃, HCl and HF; Milestone ultraWAVE microwave digestion system, 240 °C, 35 bars). Quality assurance and control protocols mainly involved using blank samples, triplicate samples, and soil standard reference materials. Approximately 5.00% of the measured samples were triplicate samples, and their analysis showed relative standard deviations were consistently below 5.00% for As, Cd, Cu, Mn, Ba, Pb, Se, Sr, Ti, and Zn. Recovery rates for HMs in the standard reference materials were above 92.60%.

Soil samples were dried at 40 °C for 48 hours, ground using an agate mortar and pestle, and sieved through a 180 µm mesh. Samples were prepared in duplicate in microplates and analyzed using an Alpha FTIR spectrometer (Bruker Optics GmbH, Ettlingen, Germany) at 4 cm⁻¹ spectral resolution over the range 4000–400 cm⁻¹. Spectra were converted to absorbance using OPUS software and averaged for each sample.

Raw spectra were preprocessed using first derivative (DV1), second derivative (DV2), or DV1 combined with Standard Normal Variate (SNV). DV1 and DV2 were applied to enhance spectral resolution and remove baseline drift, while SNV minimized particle size and light scattering effects. The spectral range 1544–600 cm⁻¹ was excluded due to high noise and water vapor interference that degraded model performance. The optimal preprocessing method for each element was selected based on minimizing the Root Mean Square Error of Cross-Validation (RMSECV) and maximizing the Coefficient of Determination (R²) during preliminary testing.

Partial Least Squares Regression (PLSR) was employed to develop calibration models relating preprocessed spectra to ICP-OES measurements. PLSR was selected because it effectively handles multicollinearity and high dimensionality inherent in spectral datasets. Models were optimized using 10-fold cross-validation, where the 67 samples were divided into 10 folds with 9 folds used for training and 1 for validation per iteration. The number of latent variables varied from 1 to 15, and the optimal number (typically 4–7 LVs) was selected by minimizing the average RMSECV across all folds. Model performance was quantitatively assessed using three standard statistical metrics: the Coefficient of Determination (R², Equation 1), the Root Mean Square Error of Cross-Validation (RMSECV, Equation 2), and the Residual Prediction Deviation (RPD, Equation 3):

Where SSE is the sum of squared errors, SST is the total sum of squares, RPD is the residual prediction deviation (with SD the standard deviation and SEP the standard error of prediction), n represents the number of samples, yi is the measured PTE concentration and y^i is the predicted concentration.

Soil pollution indexing serves as a quantitative framework for assessing the cumulative toxicity of individual PTEs and their aggregate impact on soil quality. In this study, the PTE Pollution Index (PI) was calculated using Equations 4, 5, 6.

where M_i is the monitored value of the i^th PTE, I_i is the ideal value of i^th PTE, and Si is the standard permissible value of the i^th PTE in mg·kg⁻¹. The PI assigns a single value to the overall soil quality; a higher value indicates a greater risk of harmful health effects. Generally, the critical PTE pollution index value is taken as 100. Based on the classification by (34), PI values are categorized as low (<19), medium (19–38) and high (>38).

Table 1 presents ICP-OES measured PTEs concentrations in the major soils of the three study regions. Beni Amir soils, classified as Lithic Calciustolls, showed As concentrations of 10.65 ± 1.95 mg·kg⁻¹, which fall within typical background ranges for uncontaminated agricultural soils (35, 36). When compared with WHO/FAO guidelines for agricultural soils (As: 20 mg·kg⁻¹ limit), these values remain below regulatory thresholds. These soils develop directly on limestone bedrock, creating uniform geochemical conditions (37). The moderate Sr levels (56.25 ± 44.24 mg·kg⁻¹) reflect carbonate parent material influence and remain below typical values for calcareous soils globally (38).

Kenitra soils, dominated by Typic Haplusterts, exhibited marked anthropogenic enrichment for several PTEs. Cd concentrations (3.92 ± 3.46 mg·kg⁻¹) substantially exceeded the WHO/FAO permissible limit of 3 mg·kg⁻¹. The large standard deviation (± 3.46 mg·kg⁻¹), approximating the mean, indicates the presence of localized contamination hotspots rather than uniform background levels. This heterogeneous distribution is characteristic of anthropogenic point sources, such as specific industrial activities or historical pesticide use. This accumulation occurs because smectitic clays in Vertisols have high cation exchange capacity and enhanced metal retention ability (39, 40). Zn concentrations (59.64 ± 48.83 mg·kg⁻¹) approached the upper limit of normal background levels for agricultural soils (36), though remaining below the WHO/FAO threshold of 300 mg·kg⁻¹.The similarly high variability in Zn (SD ± 48.83 mg·kg⁻¹) indicates heterogeneous contamination patterns, likely due to industrial activities combined with clay mineral retention processes. The wide range of Cd values (0.01-9.95 mg·kg⁻¹) indicates heterogeneous contamination typical of anthropogenic sources. Cu and Pb concentrations (15.40 ± 10.64 and 9.59 ± 4.61 mg·kg⁻¹, respectively) remained within typical background ranges for agricultural soils (41).

El Jadida soils, classified as Typic Calciustolls, showed unique geochemical patterns. As concentrations (17.57 ± 6.19 mg·kg⁻¹) were elevated compared to Beni Amir but remained below WHO/FAO limits. Sr concentrations (535.14 ± 272.49 mg·kg⁻¹) were notably elevated, reflecting strong marine carbonate influence and Sr²⁺ incorporation into pedogenic carbonates typical of Mollisols. Cd levels (0.95 ± 0.23 mg·kg⁻¹) exceeded typical agricultural soil backgrounds but remained lower than Kenitra values (41). The extreme Sr variability, with concentrations ranging from 116.86 to 965.54 mg·kg⁻¹, demonstrates strong geological control on element distribution through carbonate-related processes (38).

The cross-validation results for predicting potentially toxic elements reveal varying degrees of accuracy across different metals, as evidenced by the performance metrics for each element (Figure 2). Models for Ba, Zn, and Sr demonstrated the highest predictive performance, displaying R² values of 0.95, 0.94, and 0.94, respectively, indicating that these models explain over 90% of the variance in the concentration data. The slopes for Ba (0.96), Zn (0.96), and Sr (0.95) approach unity, reflecting minimal systematic error in terms of under- or overestimation. The low RMSECV values (24.94 mg.kg-1 for Ba, 8.64 mg.kg-1 for Zn, and 55.43 mg.kg-1 for Sr) further attest to the high precision of these models, making them highly reliable for accurately estimating concentrations in unknown samples from the Beni Amir, El Jadida, and Kenitra regions.

High predictive accuracy for Zn has been reported in previous studies, where it was attributed to strong interactions with soil components such as clay and humic substances (42, 43). The excellent performance of Ba aligns with earlier research indicating that alkaline earth elements might display improved spectral sensitivity as a result of their interactions with soil carbonates (42). Similarly, Sr’s strong performance can be attributed to its association with carbonate minerals common in the sedimentary formations of the study regions.

The models for Mn and Cu demonstrate strong predictive capabilities, with R² values of 0.90 for both elements, slopes approaching unity (0.93 and 0.89, respectively), and relatively low RMSECV values (55.04 mg.kg-1 for Mn and 2.63 mg.kg-1 for Cu), indicating that the models can capture concentration trends effectively with minimal error. These results contrast favorably with previous studies that documented lower predictive performance for similar elements. For example, A study reported R² values of only 0.50, 0.42, and 0.55 for Zn, Mn, and Cu, respectively, in the validation set, indicating poor predictive performance (44). Another one, reported coefficients of determination of 0.76, 0.78, and 0.64 for Mn, Zn, and Cu, respectively, suggesting that PLSR models may have limited predictive power for estimating metal concentrations in some soil types (45). The close agreement between standard deviations of ICP-OES and FTIR-predicted data (e.g., Zn: 48.82 vs. 46.72 mg·kg⁻¹; Cd: 3.45 vs. 2.95 mg·kg⁻¹ in Kenitra soils) demonstrates that FTIR-PLSR captures both central tendency and spatial heterogeneity, including the localized contamination hotspots identified in Section 3.1, which is essential for accurate environmental risk assessment.

Cd, As, and Se exhibited moderate predictive performance with R² values of 0.84, 0.82, and 0.82, respectively, with some degree of underestimation as indicated by their slopes being less than unity (0.88, 0.88, and 0.84). This performance is consistent with prior research, which has documented challenges in predicting elements with weaker or more variable spectral features. Similar trends have been observed in previous studies, particularly for As and Se, which often display weak and overlapping absorption bands in the MIR range (43). The performance for As (R² = 0.82) aligns with earlier research by (46), who demonstrated PLSR prediction performance for As with an R² of 0.69.

The models for Ti and Pb exhibit moderate to low predictive accuracy. For Pb, the R² value is 0.48, indicating that less than half of the variability in lead concentration is captured by the model, which is lower than previous findings that reported approximate R² values of 0.70 (47). The slope of 0.52 suggests significant underestimation of Pb concentrations across the concentration range, while the RMSECV of 2.91 mg.kg-1 implies considerable prediction error. The model for Ti shows moderate accuracy with an R² of 0.66 and a relatively high RMSECV of 257.90 mg.kg-1, suggesting considerable prediction error. The slope for Ti (0.72) indicates systematic underestimation across the concentration range, likely due to the non-homogeneous distribution of Ti content in soils and limitations in the spectral data for this element.

The weak predictive performance of Pb and Ti is consistent with reports indicating significant challenges in using MIR spectroscopy for these elements (43). The inertness of Pb and its weak interactions with soil organic matter or minerals reduce its spectral detectability (48). Likewise, Ti’s complex mineralogical associations, especially with silicates, likely obscure their absorption features in the MIR range, complicating accurate modeling efforts.

The validation results confirm the practical applicability of MIR-FTIR and PLSR models for rapidly predicting key soil contaminants (Ba, Zn, Sr, Mn, Cu), while highlighting the need for methodological refinements or complementary analytical approaches for elements with moderate or low predictive accuracy. Elements with strong soil-mineral interactions and distinct spectral signatures showed superior predictive performance, supporting prior studies linking these properties to robust spectral responses (49). These results emphasize the importance of element-specific properties in improving predictive accuracy for soil contamination assessment in Moroccan agricultural soils. FTIR predictions accurately reproduced the soil-specific geochemical patterns observed by ICP-OES (Table 2). For Lithic Calciustolls, FTIR maintained low As predictions (10.70 ± 1.76 mg.kg⁻¹) and preserved the geochemical homogeneity. The simple mineralogy of these young soils facilitates accurate spectral interpretation (50). Typic Haplusterts presented spectroscopic challenges due to swelling clay complexity, but FTIR successfully predicted Cd enrichment (3.79 ± 2.96 mg.kg⁻¹) and Zn accumulation (58.75 ± 46.72 mg.kg⁻¹). The preservation of standard deviations between methods demonstrates that FTIR maintains natural geochemical variability (51). For Typic Calciustolls, FTIR exceptionally reproduced Sr concentrations (543.00 ± 253.59 mg.kg⁻¹ vs 535.14 mg.kg⁻¹ ICP-OES). As predictions (17.51 ± 4.30 mg.kg⁻¹) were equally accurate, reflecting organometallic complex detection in the organic-rich mollic horizons (52).

Statistical comparison revealed complete agreement between ICP-OES and FTIR methods across all soil types (Table 3). Paired t-tests of 30 comparisons showed no significant differences (all p > 0.05), achieving 100% agreement. A study demonstrated that uniform performance across contrasting soil types demonstrates method robustness independent of pedological context (53).

Mean differences were minimal across all PTEs. As showed mean differences of 0.05 mg.kg⁻¹ for Beni Amir (p = 0.83), 0.21 mg.kg⁻¹ for Kenitra (p = 0.65), and 0.06 mg.kg⁻¹ for El Jadida (p = 0.92), all remaining below 0.5 mg.kg⁻¹ across all soil types. Cu exhibited comparable precision with differences of 0.03 mg.kg⁻¹ (Beni Amir), 0.03 mg.kg⁻¹ (Kenitra), and 0.18 mg.kg⁻¹ (El Jadida), while Pb showed differences below 0.27 mg.kg⁻¹ across all regions. Cd demonstrated exceptional agreement with mean differences of 0.24 mg.kg⁻¹ (Beni Amir), 0.12 mg.kg⁻¹ (Kenitra), and 0.09 mg.kg⁻¹ (El Jadida). The consistency of these low differences across three contrasting soil types, Lithic Calciustolls with simple limestone mineralogy, Typic Haplusterts with complex smectitic clays, and Typic Calciustolls with organic-rich horizons demonstrates the reliability of FTIR predictions independent of soil matrix composition.

PTEs associated with mineral phases showed equally strong agreement. Sr differences averaged 7.86 mg.kg⁻¹ for El Jadida Calciustolls (1.5% of mean concentration of 535.14 mg.kg⁻¹), 2.97 mg.kg⁻¹ for Beni Amir, and 2.58 mg.kg⁻¹ for Kenitra, confirming accurate carbonate-associated element detection (40). This precision results from consistent carbonate spectral responses regardless of soil type, as demonstrated by (54) for various calcareous soils. Ba showed mean differences of 8.55 mg.kg⁻¹ (Beni Amir), 6.87 mg.kg⁻¹ (Kenitra), and 0.03 mg.kg⁻¹ (El Jadida), while Mn differences remained at 2.89 mg.kg⁻¹, 0.37 mg.kg⁻¹, and 5.91 mg.kg⁻¹ respectively, all below 10 mg.kg⁻¹. Zn demonstrated exceptional agreement with differences of 1.14 mg.kg⁻¹ (Beni Amir), 0.88 mg.kg⁻¹ (Kenitra), and 0.48 mg.kg⁻¹ (El Jadida), reflecting robust detection across different soil matrices (55).

P-values showed remarkable consistency between soil types: 0.65 for Lithic Calciustolls, 0.72 for Typic Haplusterts, and 0.69 for Typic Calciustolls. This uniformity demonstrates that FTIR performance is independent of soil mineralogy, texture, or chemistry (56). Statistical equivalence confirms model transferability between soil types, essential for large-scale applications (57).

This validation across three contrasting soil types establishes FTIR robustness for PTEs analysis independent of pedological context. The analytical equivalence between ICP-OES and FTIR enables routine environmental monitoring applications. FTIR’s ability to preserve soil-specific geochemical signatures while maintaining accuracy across different pedological contexts represents a significant methodological advance for soil quality assessment in Mediterranean environments (5).

The correlation analysis showed that FTIR spectroscopy maintains the same geochemical relationships observed in ICP-OES measurements (Figure 3), with correlation differences (Δr) below 0.083 across all property-metal pairs, proving that FTIR predictions capture real soil chemical processes rather than false statistical relationships.

pH was the most important soil property, showing very strong negative correlations with strontium concentrations (r = -0.833 for ICP-OES vs r = -0.834 for FTIR), reflecting pH-dependent carbonate precipitation mechanisms that reduce Sr²⁺ availability under alkaline conditions. FTIR spectroscopy monitors these processes through carbonate deformation vibrations at 863 cm⁻¹, which are sensitive to strontium incorporation (58). As showed significant negative correlations with pH (r = -0.580 ICP-OES, r = -0.626 FTIR), consistent with enhanced As (V) mobility at elevated pH through surface charge changes on iron and aluminum oxides, leading to increased As extractability (59). In contrast, Se showed positive pH correlations (r = 0.345 ICP-OES, r = 0.383 FTIR), reflecting selenate stability under alkaline conditions and different adsorption mechanisms compared to arsenate.

CEC showed strong negative correlations with As (r = -0.531 ICP-OES, r = -0.614 FTIR) and strontium (r = -0.691 ICP-OES, r = -0.694 FTIR), reflecting competitive adsorption mechanisms where high CEC values indicate many permanent negative charges that preferentially bind major cations (Ca²⁺, Mg²⁺), reducing available sites for trace metal retention through Coulombic exclusion effects. FTIR detects these processes through clay mineral OH stretching vibrations at 3564 cm⁻¹, as metal substitution changes hydrogen bonding environments a round exchange sites, causing measurable frequency shifts orc (60, 61). Carbonate content showed moderate negative correlations with strontium (r = -0.442 ICP-OES, r = -0.444 FTIR), reflecting the complex relationship between mineral precipitation and metal extractability, where Sr²⁺ incorporation into carbonate structures removes strontium from easily extractable pools despite its affinity for carbonate minerals (58). TOC showed moderate negative correlations with strontium (r = -0.435 ICP-OES, r = -0.443 FTIR), reflecting metal-organic complexation through carboxyl and phenolic functional groups that form stable chelate complexes reducing metal bioavailability. FTIR monitors these interactions through C=O stretching at 1726 cm⁻¹ and C-H stretching at 2888 cm⁻¹ (15).

The multi-mechanism detection capability of FTIR spectroscopy, including cation exchange monitoring (3564 cm⁻¹), carbonate interactions (863 cm⁻¹), organic complexation (1726 cm⁻¹), and silicate incorporation (1002 cm⁻¹), explains the exceptional agreement between ICP-OES and FTIR measurements across all soil property-metal relationships (62). This mechanistic validation demonstrates that FTIR spectroscopy directly monitors the molecular environments where PTEs-soil interactions occur, rather than relying on extraction procedures that may alter metal speciation during analysis. The preservation of fundamental geochemical relationships in FTIR predictions provides compelling evidence that spectroscopic methods capture actual soil chemical processes governing PTEs behavior, positioning FTIR as a reliable alternative for rapid PTEs assessment that maintains geochemical understanding essential for environmental risk evaluation (63).

The PTE Pollution Index (PI) was calculated for all 67 soil samples using both ICP-OES measured concentrations and FTIR-predicted values to evaluate the practical applicability of spectroscopic methods for environmental risk assessment (Table 4). The PI analysis demonstrated exceptional agreement between methods across all soil types. Mean PI values were nearly identical: 1.44 ± 1.10 (ICP-OES) vs. 1.57 ± 1.02 (FTIR) for Beni Amir, 2.35 ± 1.16 vs. 2.27 ± 0.72 for Kenitra, and 0.61 ± 0.33 vs. 0.56 ± 0.40 for El Jadida, with overall means of 1.54 for both methods (difference = 0.00). Regional variability in PI values reflects differences in soil geochemistry and parent material composition. All samples from all regions were classified as “Low” pollution status (PI < 19) by both methods.

Statistical validation confirmed method equivalence. Paired t-tests comparing ICP-OES and FTIR within each region detected no significant differences (Beni Amir: p = 0.059; Kenitra: p = 0.658; El Jadida: p = 0.668). One-way ANOVA comparing all ICP-OES versus FTIR samples revealed no significant method effect (F = 0.30, p = 0.586, Table 5), demonstrating that the choice of analytical method contributes no systematic variation to observed PI values. This statistical equivalence validates FTIR for pollution indexation despite fundamental differences in measurement principles.

These findings align with recent spectroscopic validation studies in agricultural soils. (64) Demonstrated strong correlations (r > 0.80) between spaceborne hyperspectral imaging and ICP-MS measurements for nickel concentration in agricultural soil, while (65) validated hyperspectral prediction of ecological risk indices with strong correlations for Cd (r = 0.95) in rice paddy soils. (66) successfully applied VIS-NIR spectroscopy for predicting contamination and ecological risk indices in agricultural soils under anthropogenic pressure. (57) demonstrated that field-portable spectrometers enable in-situ soil analysis with minimal sample preparation, providing rapid field-scale analysis that significantly exceeds laboratory ICP throughput. This advantage is particularly relevant for resource-constrained settings and regional assessment programs requiring analysis of numerous samples.

In conclusion, the combination of highly similar descriptive statistics, non-significant paired t-tests and one-way ANOVA (F = 0.30, p = 0.586) provides compelling evidence for FTIR-ICP-OES equivalence in pollution index calculation for agricultural soils. The universally low PI classification indicates that soils in the study regions remain within acceptable contamination limits. This validation supports implementation of FTIR-based screening for environmental risk assessment, enabling more extensive and cost-effective soil monitoring.