Integration of Slurry–Total Reflection X-ray Fluorescence and Machine Learning for Monitoring Arsenic and Lead Contamination: Case Study in Itata Valley Agricultural Soils, Chile

Abstract

The accuracy of determining arsenic and lead using the optical technique Slurry–Total Reflection X-ray Fluorescence (Slurry-TXRF) was significantly enhanced through the application of a machine learning method, aimed at improving the ecological risk assessment of agricultural soils. The overlapping of the arsenic Kα signal at 10.55 keV with the lead Lα signal at 10.54 keV due to the relatively low resolution of TXRF could compromise the determination of lead. However, by applying a Partial Least Squares (PLS) machine learning algorithm, we mitigated interference variations, resulting in improved selectivity and accuracy. Specifically, the average percentage error was reduced from 15.6% to 9.4% for arsenic (RMSEP improved from 5.6 mg kg⁻¹ to 3.3 mg kg⁻¹) and from 18.9% to 6.8% for lead (RMSEP improved from 12.3 mg kg⁻¹ to 5.03 mg kg⁻¹) compared to the previous univariable model. This enhanced predictive accuracy, within the set of samples concentration range, is attributable to the efficiency of the multivariate calibration first-order advantage in quantifying the presence of interferents. The evaluation of X-ray fluorescence emission signals for 26 different synthetic calibration mixtures confirmed these improvements, overcoming spectral interferences. Additionally, the application of these models enabled the quantification of arsenic and lead in soils from a viticultural subregion of Chile, facilitating the estimation of ecological risk indices in a fast and reliable manner. The results indicate that the contamination level of these soils with arsenic and lead ranges from moderate to considerable.

1. Introduction

In environmental geochemistry, it is crucial to understand the concepts of geological baselines and critical levels. This understanding is important for differentiating between naturally occurring substance levels and pollution caused by human activities [1]. Several ecological indicators based on these geological baselines have been suggested to evaluate the extent and impact of soil contamination by trace elements, including the Geoaccumulation Index (GEOI), Enrichment Factor (EF), and Contamination Factor (CF) [2].

Establishing the baseline levels of trace elements and ecological indices in a particular location is important for detecting contamination. Still, it requires the systematic gathering and examination of a large number of samples [3].

Quantitative analyses in soil studies often utilize various optical analytical methods, including traditional approaches like inductively coupled plasma mass spectrometry (ICP-MS) [4], flame atomic absorption spectroscopy (FAAS) [5], inductively coupled plasma emission spectroscopy (ICP-OES) [6], graphite furnace atomic absorption spectroscopy (GF-AAS) [7], and cold vapor atomic absorption spectrometry (CV-AAS) [8].

Total X-ray fluorescence spectroscopy (TXRF) is a reliable and convenient alternative method for determining trace element concentrations. This versatile technique can be applied with or without sample digestion, making it a highly favored choice [2,9,10].

However, these methods often lead to significant spectral interferences. Studies have been carried out to explore the use of machine learning algorithms such as Partial Least Squares Regression (PLS-R) in liquid matrices [11], the Synchrotron Radiation X-ray Fluorescence technique [12], and Cd-K in soils by TXRF [13], where these models improve the accuracy of quantification in the presence of spectral interferences.

Several studies have examined the levels of heavy metals or potentially toxic elements in areas primarily used for agriculture and mining [2,14,15,16,17,18,19,20,21]. However, there is a lack of studies on effectively monitoring arsenic and lead contamination in agricultural soils, particularly in regions like Itata Valley. While there are standardized methods suggested by international quality standards for the determination of arsenic and lead [22], these methods are often slow, expensive, and environmentally unfriendly.

Therefore, it is crucial to continue developing and improving analytical methods for trace element analysis in soil samples, as well as exploring innovative techniques and approaches such as machine learning and X-ray fluorescence to overcome challenges and improve quantification accuracy in the context of low-resolution analytical methods. [23].

This study aimed to assess the effectiveness of the specific method of slurry-TXRF combined with machine learning models in analyzing arsenic and lead levels in viticultural soils. This method offers a rapid and efficient way to estimate and monitor contamination levels using various ecological indices, providing an innovative alternative for gaining valuable insights into the extent and potential risks of heavy-metal contamination in vineyard soils. These insights can help make informed decisions about soil management practices and ensure sustainable agriculture for the quality and safety of products.

2. Materials and Methods

This study on improving lead and arsenic measurement using slurry-TXRF and machine learning is part of a larger research effort focused on studying heavy-metal contamination in agricultural soils. Medina-Gonzalez et al.’s (2023) [2] study served as the basis for our methodology with certain adaptations.

2.1. Study Area and Samples

A well-known wine-producing region in southern Chile was chosen as the focus of this study. This region covers around 311,418 hectares, with approximately 13,030 hectares allocated to agricultural activities, specifically for wine production [24]. The Itata Valley is notable for its cultivation of grape varieties like País, Muscat of Alexandria, and Cinsault, with the latter being the most widely grown in terms of cultivation area [25].

Geographically, the Itata Valley is part of the Central Valley and Coastal Range, with elevations not exceeding 400 m above sea level. The area benefits from a Mediterranean climate, with distinct rainy and dry seasons, influenced by the Itata River, which flows from east to west across the valley. The presence of the Coastal Range creates climatic variations, with the western slopes receiving more rainfall compared to the drier eastern slopes [24].

The valley’s soil originates primarily from three parent materials:

• Eroded Metamorphic Rock: Found in higher elevations with steep or complex slopes, this soil is characterized by shale, sandstone, phyllite, and slates. It has a clayey texture with slow water infiltration and is prone to forming Catena due to topography and drainage patterns.
• Granite Origin: Derived from granite and diorite rocks, this soil also has a clayey texture and low water infiltration, making it susceptible to erosion. It is typically found in hilly areas with complex slopes.
• Fine Alluvial Sediment: Formed from fluvio-glacial sediments transported by rivers, this soil has a loamy–clayey texture and poor drainage. The thickness of these sediments varies significantly across the valley.

In summary, the soils in the Itata Valley are primarily derived from eroded metamorphic rocks, granite and diorite, and fine alluvial sediments, each contributing to unique characteristics such as soil texture, water infiltration, and erosion susceptibility [25].

The soil sampling was carried out in two phases, namely June and October of 2023. These samples were obtained from five vineyards (A, B, C, D, E) at a total of 48 sampling points within the area corresponding to the Itata Valley, following the owners’ authorization and using the combined method as recommended by Servicio Agrícola y Ganadero de Chile [26] (Figure 1).

Approximately two kilograms of the sample were collected at each sampling location by taking five subsamples from the corners and center of a 1 × 1 m grid to create a composite sample. The topsoil samples (0–20 cm depth) were collected while removing any plant cover present and deep soil samples at a depth of 150 cm using a Dutch auger. Surface soils are most affected by human activities, whereas deep soils represent relatively undisturbed samples considered free from contamination [27].

The samples were packed in polyethylene bags for transport to the laboratory and then dried in an oven at 40 °C for 48 h. Next, each sample was sieved using a nylon sieve to remove coarse material and other residues, leaving only the fine material (<2 mm). The sieved samples were stored in tightly sealed plastic bags for further analysis [28].

2.2. Sample Preparation

2.2.1. PLS Model Samples

Standard calibration and external validation samples were prepared using 1000 mg L⁻¹ solutions of As and Pb. Ultra-pure deionized water with a resistance of 18.2 MΩ cm⁻¹ obtained from a Milli-Q quality purification system was used for dilution purposes. A silicone solution in isopropanol was applied to create a silicon film on reflective disks of quartz glass before the deposition of the sample. Working solutions with concentrations of 100 mg L⁻¹ were utilized to generate a total of 26 synthetic calibration solutions (

Table 1. Composition of synthetic calibration mixtures.

Sample	As (mg kg−1)	Pb (mg kg−1)
C1	10	34
C2	12	72
C3	48	13
C4	42	20
C5	37	67
C6	33	15
C7	33	19
C8	32	49
C9	8	39
C10	43	65
C11	35	36
C12	49	5
C13	10	4
C14	45	78
C15	13	31
C16	33	69
C17	41	36
C18	24	22
C19	4	36
C20	7	72
C21	28	23
C22	33	36
C23	42	18
C24	36	47
C25	0	43
C26	12	27
Range	49	74
Minimum	0	4
Maximum	49	78

) and 8 external validation samples (

Table 2. Composition of synthetic mixtures for external validation.

Sample	As (mg kg−1)	Pb (mg kg−1)
V1	36	6
V2	19	54
V3	1	32
V4	12	51
V5	33	78
V6	16	42
V7	21	71
V8	21	7

All samples contained 10 ng of germanium as internal standard.

) for these elements in 2 mL Eppendorf tubes.

Each sample preparation included a random mixing of concentrations to avoid any potential correlations that could affect the calibration process. The concentration ranges were intentionally chosen to facilitate the measurement of arsenic and lead in environmental samples with varying levels of contamination. Sample preparation involved using the drop-drying method, which entails applying a small volume (such as 10 μL) of the mixture onto 30 mm diameter quartz sample disks and drying them with an infrared lamp.

The resulting solid residues on the sample holders were then analyzed in an S4TStar TXRF spectrometer after the evaporation of the liquid phase [2].

To create a model in milligrams per kilogram (mg kg⁻¹) concentration units from liquid calibration samples, enabling us to evaluate its suitability in solid matrices, conversions in concentrations were conducted using Equation (1):

$$\left[{\mathrm{mgkg}}^{-1}\right]={\displaystyle \frac{\left[{\mathrm{mgL}}^{-1}\right]{\times V}_{\mathit{dep}}}{S_{\mathit{dep}}}}$$

(1)

where V_dep represents the volume of liquid calibration sample deposited in liters, while S_dep denotes the quantity of dry solid sample deposited in kilograms, based on the specific methodologies employed.

2.2.2. Slurry Sample Preparation

The soil sludge samples were prepared in sterilized 2.5 mL Eppendorf vials by combining 30 mg of dried and pulverized soil with 1500 μL of Triton-X as a surfactant using an agate mortar. The mixtures were homogenized for 300 s with an electronic shaker. After this, 10 μL of the thoroughly homogenized suspension was extracted and placed into the TXRF’s 30 mm diameter quartz sample holder immediately to prevent potential sedimentation. A silicone solution in isopropanol was then applied to create a silicon film on the quartz glass disk before depositing the sample, which was followed by drying it with an infrared lamp. To evaluate accuracy and suitability, loamy clay certified reference material for trace metals from Sigma-Aldrich was used to test the model’s performance in soil samples.

2.3. TXRF Spectra and Data Acquisition

This study used a TXRF S4TSTAR analytical system (Bruker^® AXS Microanalysis GmbH, Berlin, Germany) for fluorescent emission measurements. The X-ray source operated at 50 W with an electron flow of 1 mA at 50 kV and was cooled by air. The system included a multi-layer carbon–nickel monochromator of 150 nm and a high-resolution silicon semiconductor detector XFlashR with an active area of 100 mm2 and energy resolution < 140 eV. The specific range of interest was defined from 10.27 keV to 12.78 keV, focused on the overlap of the arsenic Kα line (10.54 keV) and lead Lα line (10.55 keV). Counts were recorded using the SPECTRA 7.5.0.3 software for data analysis [29], generating a .txt format file processed through PIROUETTE 4.5 software [30].

2.4. Quantification Methods

For the aim of assessing the efficacy of the PLS model in quantifying arsenic and lead, in comparing its performance to classical deconvolution methods utilizing the SPECTRA 7.5.0.3 software, specifically the deconvolution routine (SuperBayes), the quantification was carried out employing internal standardization with germanium (10 ng).

A set of 26 synthetic mixtures served as the basis for PLS analysis. These mixtures were deliberately correlated with their known concentrations and employed as calibration samples. The fluorescence spectra within the calibration set underwent preprocessing. Specifically, mean centering was applied as the sole transformation method without additional alterations.

Latent variables were chosen through the cross-validation procedure employing the leaving-one-out method. This method systematically excluded each sample from the calibration set one at a time, utilizing it as a prediction object. With this, the root mean square error of cross-validation (RMSECV), calculated as per Equation (2), served as the criterion to determine the optimal number of latent variables essential for establishing the calibration model.

To assess the effectiveness of the Partial Least Squares model in quantifying arsenic and lead, its performance was compared to those of classical deconvolution methods using Spectra 7.5.0.3 software’s SuperBayes routine with internal standardization using germanium (10 ng). A set of 26 synthetic mixtures correlated with known concentrations served as calibration samples for PLS analysis. These fluorescence spectra underwent mean-centering preprocessing without any further transformation. The optimal number of latent variables required for establishing the calibration model was determined through cross-validation and root mean square error of cross-validation [31], calculated according to Equation (2):

$$\mathrm{RMSECV}=\sqrt{\sum {\left(y_i-{\hat{y}}_i\right)}^2/n}$$

(2)

where ŷ_i represents the estimated concentration values obtained through cross-validation, with each sample i being estimated using a model constructed from the sample set excluding that particular sample. Here, y_i denotes the known concentration variable value, and n denotes the total number of samples in the calibration set.

The same sample set used for the direct quantification of arsenic and lead through deconvolution (Table 2) using the SPECTRA software also played another important role: external validation of the model. External validation is a crucial practice in chemometrics that helps guard against overfitting in calibration models, ensuring their robustness and generalizability [32].

The root mean square error of prediction (RMSEP) was used to evaluate the predictive accuracy of the multivariate model [33]. It is calculated similarly to RMSECV, but in this case, each predicted sample (y_i) from the final model is compared with the corresponding reference value from an external validation set (ŷ_(i,ref)), where N represents the size of the external validation set. The calculation for the RMSEP follows Equation (3):

$$\mathrm{RMSEP}=\sqrt{\sum {\left(y_{\mathrm{i}}-{\hat{y}}_{i,\mathit{ref}}\right)}^2/N}$$

(3)

Linear regression can identify biases in the predictions of external samples, similar to its use in internal validation procedures. Equation (4) shows the linear regression approach used to assess systematic biases within the predictions for external samples.

$$\hat{\mathrm{y}}=a+yb$$

(4)

The comparison between predicted values ($\hat{\mathrm{y}}$) and reference values (y) is crucial for evaluation. Ideally, predictions made by the calibration model should exhibit a linear relationship with known values—a line with a slope (b) of 1 and an intercept (a) of 0. Any deviations from this linear behavior indicate issues arising during the analytical procedure [34]. The Elliptical Joint Confidence Region (EJCR) method was employed to assess the correlation of results and obtain a statistically significant confidence level of 95% [35].

Analytical Figures of Merit (AFOM’s)

Figures of Merit for the PLS approach were calculated according to the NAS theory. This concept works by isolating relevant signals associated with a particular component of interest from the various interfering elements present in the spectra. The specific method utilized in this research follows the approach described by Short et al. [36], concentrating on deriving figures of merit using the NAS theory applied to spectral data.

2.5. Background Values and Ecological Indices

The assessment of natural or baseline values required the examination of data from deep soil samples following the procedures outlined in ISO 19258 [37]. There is no standardized method for setting threshold values to identify samples with abnormally high levels of element concentrations, as noted by Reimann et al. [38]. Nonetheless, this research applies the “MAD” technique, which makes use of the range from Median ± 2MAD (median absolute deviation) to pinpoint values suggestive of elevated element concentrations.

Various environmental indicators are frequently utilized to evaluate possible soil pollution, aiding in the differentiation between natural trace element accumulation and that caused by human actions [18,39]. One commonly used method is the Geoaccumulation Index, which is determined using a specific equation (Equation (5)):

$${\mathrm{GEOI}=\log }_2(C_i/1.5C_b)$$

(5)

where C_i is the measured content of the element in the soil, and C_b is the estimated background value of that element in the soil. According to Muller [40], the GEOI for a particular element is calculated and then classified as shown in

Table 3. Contamination level according to the Geoaccumulation Index [18].

GEOI Group	Level of Contamination
GEOI ≤ 0	Uncontaminated
0 < GEOI ≤ 1	Slightly contaminated
1 < GEOI ≤ 2	Moderately contaminated
2 < GEOI ≤ 3	Moderately to heavily contaminated
3 < GEOI ≤ 4	Heavily contaminated
4 < GEOI ≤ 5	Heavily to extremely contaminated
GEOI > 5	Extremely contaminated

.

The Enrichment Factor (EF) is determined by standardizing the concentration of a measured element with respect to a reference element. The selection of the reference element relies on its minimal variability in occurrence within the specific environment [38] and its independence from anthropogenic influences [41]. Commonly used reference elements comprise Sc, Mn, Ti, Al, and Fe. This research chose Fe as the reference element due to its significantly low variation in viticultural soils in the Itata Valley [1]. Its abundance in the Earth’s crust and its connection with the soil matrix validate its choice as a stable reference element [42]. The EF is calculated using Equation (6):

$${\mathrm{EF}=(C}_i/{\mathit{CFe}}_i)/{(C}_b/{\mathit{CFe}}_b)$$

(6)

where C_i represents the concentration of the particular element, CFe_i denotes the quantity of Fe as a reference constituent, C_b is the reference content derived via the MAD method, while CFe_b signifies background content for Fe. Enrichment Factor values are divided into five categories based on

Table 4. Contamination level according to the Enrichment Factor (EF) [18].

EF Group	Level of Contamination
EF < 2	Deficiency to minimal enrichment
2 < EF < 5	Moderate enrichment
5 < EF < 20	Significant enrichment
20 < EF < 40	Very high enrichment
EF > 40	Extremely high enrichment

:

The Contamination Factor (CF) is calculated using Equation (7):

$$\mathrm{CF}=C_i/C_b$$

(7)

where C_i represents the concentration of the examined element in the soil, and C_b is the content of background values obtained using the MAD method. Hakanson [43] defines the categories for CF as presented in

Table 5. Contamination level according to Contamination Factor (CF) [18].

CF Group	Level of Contamination
CF < 1	Low contamination
1 < CF < 3	Moderate contamination
3 < CF < 6	Considerable contamination
CF > 6	Very high contamination

.

3. Results

3.1. Classical Regression: Deconvolution

Table 6. Comparison of As and Pb quantification results by Internal Standardization vs. PLS.

Sample	Concentration (mg kg−1)		Internal Standardization (mg kg−1)		Relative Bias (%)		PLS (mg kg−1)		Relative Bias (%)
	As	Pb	As	Pb	As	Pb	As	Pb	As	Pb
A	36	6	45.3	8.1	25.8	35.0	39.8	7.2	10.6	20.0
B	19	54	24.2	68.2	27.4	26.3	15.2	62.4	−20.0	15.6
C	1	32	2.5	33.9	150.0	5.9	1.9	31.4	90.0	−1.9
D	12	51	16.2	60.3	35.0	18.2	9.3	56.6	−22.5	11.0
E	33	80	40.3	87.6	22.1	9.5	38.5	83.4	16.7	4.3
F	16	42	18.5	52.2	15.6	24.3	18.2	37.4	13.8	−11.0
G	21	71	28.3	98.1	34.8	38.2	17.3	64.2	−17.6	−9.6
H	21	7	23.5	12.2	11.9	74.3	22.3	10.8	6.2	54.3
RMSEP (mg kg−1)			5.6	12.3			3.3	5.0
RMSEP (%)			15.6	18.9			9.4	6.8

shows the results obtained for the quantification of arsenic and lead in conditions where there is overlap, using the classical method of internal standardization followed by signal deconvolution.

The results indicate significant positive biases for both arsenic and lead, with relative biases ranging from 11.9% to 150% for arsenic and from 5.9% to 74.3% for lead. These values, respectively, in relation to the reference values of the external validation set, reveal a noticeable systematic error. The calculated RMSEP for this quantification method is 5.6 mg kg⁻¹ for arsenic and 12.3 mg kg⁻¹ for lead, resulting in an average percentage error of 15.6% and 18.9%, respectively, within the concentration range of the sample set.

3.2. Machine Learning Regression: Partial Least Squares

The X-ray fluorescent emission spectra of the 26 samples, which were utilized to develop the PLS models, cover a range of energies, including the Kα (10.54 keV) and Kβ (11.73 keV) lines for arsenic, as well as Lα (10.55 keV) and Lβ (12.61 keV) lines for lead, as shown in Figure 2.

When considering two latent variables, the model shows RMSECVs of 2.95 mg kg⁻¹ and 3.92 mg kg⁻¹. For external validation, there is an RMSEP of 3.31 mg kg⁻¹ and 5.03 mg kg⁻¹ for As and Pb, respectively. The discrepancy between both parameters is expected due to the tendency of cross-validation to be overly optimistic in comparison to external sample validation.

Additionally, it is observed that both cross-validation and external validation predictions for these elements exhibit a high level of correlation with the real concentrations of the samples, achieving correlation coefficients rVal = 0.981 and rPred = 0.972 for As and rVal = 0.986 and rPred = 0.982 for Pb (Figure 3).

The quantification biases for arsenic and lead, observed in both multivariate models and classical quantification methods (internal standardization and deconvolution), have significantly reduced. There is no longer a pronounced positive trend, and the percentage of biases has notably decreased. The summary of all these parameters for the PLS model is presented in

Table 7. Characteristics and parameters of the PLS model for As and Pb.

	As Model	Pb Model
RMSECV (mg kg−1)	2.95	3.92
rVal	0.981	0.986
RMSEP (mg kg−1)	3.31	5.03
rPred	0.972	0.982
Calibration Samples	26
Number of Variables	257
Pretreatment	Mean-centering
Latent Variables	2
Range [As] (mg kg−1)	0 to 49
Range [Pb] (mg kg−1)	4 to 78

.

When examining the regression vectors depicted in Figure 4, it can be observed that the multivariate model’s predictions for arsenic concentration predominantly rely on signals produced within the energy range corresponding to its primary emission line (Kα). Signals from the secondary emission line (Kβ) also have a lesser impact. This finding highlights how incorporating multiple variables and assessing their contributions improves the accuracy of predicting arsenic levels, effectively managing interference from lead’s fluorescent signal.

In contrast, there is a more evident overlap of signals, particularly from lead’s most intense emission line (Lα). Consequently, signals from its weaker line (Lβ) become more significant within the model. This indicates that reliance on signals from the Lβ zone is higher due to interference with the stronger Lα line.

3.3. Elliptical Joint Confidence Region (EJCR)

In employing the EJCR correlation method at a significance level (α) of 0.05, it has been statistically shown that determinations made using univariate methods like internal standardization and deconvolution lack accuracy. This conclusion is drawn from the failure to meet the criteria of a slope of 1 and an intercept of 0, indicating a deviation from the ideal correlation.

Figure 5 visually illustrates this difference. The ellipse in the figure demonstrates the disparity between the ideal point representing slope vs. intercept and the observed correlation, highlighting that there is a significant distance between the observed data points and their ideal position.

The deconvolution via software prior internal standardization fails to satisfy the EJCR criteria, as evidenced by significant deviations in slope and intercept from their ideal values (α = 0.05).

In contrast, multivariate calibration using Partial Least Squares demonstrates accuracy in quantification. Adhering to the EJCR criteria with an ideality point within the ellipse confirms that both slope and intercept statistically approach 1 and 0, respectively (α = 0.05).

This comparison emphasizes how effective multivariate calibration through PLS is in addressing overlapping X-ray fluorescence signals of arsenic and lead. Consequently, this method produces more precise and accurate results than deconvolution processes.

3.4. AFOM

Table 8. Analytical figures of merit using internal standardization and PLS methods are presented, including limit of detection (LOD), limit of quantification (LOQ), sensitivity, and analytical sensitivity.

Figure of Merit	Internal Standard		PLS
	As	Pb	As	Pb
LOD (mg kg−1)	5.42 × 10−2	6.28 × 10−2	2.56 × 10−4	3.48 × 10−4
LOQ (mg kg−1)	1.64 × 10−1	1.90 × 10−1	7.69 × 10−4	1.05 × 10−3
Sensitivity (counts (mg kg−1)−1)	432	298	474	402
Analytical Sensitivity ((mg kg−1)−1)	18.3	12.7	59.6	63.3

displays the analytical performance metrics for each quantification approach. It is evident from the table that the PLS method has resulted in an enhanced limit of detection by two orders of magnitude and limit of quantification by three orders and two orders of magnitude for As and Pb, respectively, compared to internal standard calibration. The analytical sensitivities obtained through PLS models have also seen important improvements.

3.5. Application in Soil Samples

Due to the lack of noticeable differences in the spectra between liquid and well-prepared solid samples within the relevant range, a comparison of techniques was conducted using a validated soil sample (Clay 2—CRM051). The assessment employing a certified soil sample, based on the t-test criteria (α = 0.05, n = 3), reveals that both techniques yield results that are not statistically different from the certified value (

Table 9. Comparison of accuracy for arsenic (As) and lead (Pb) quantification in Certified Soil Sample Clay 2—CRM051 using internal standardization and PLS methods.

	Certified Value		Internal Standard (n = 3)		PLS Model (n = 3)
	As	Pb	As	Pb	As	Pb
Mean (mg kg−1)	45.5	68.1	55.3	79.2	51.2	63.2
SD 1	4.45	7.6	6.23	11.2	7.23	8.23
RSD 2 (%)	9.78	11.16	11.27	14.14	14.12	13.02
p (α = 0.05)	-	-	0.077	0.215	0.31	0.483

¹ Standard deviation; ² relative standard deviation.

). However, in terms of bias and accuracy, the PLS model surpasses the quantification achieved through internal standardization and signal deconvolution. The PLS model demonstrates lower bias and higher accuracy compared to other methods. Subsequently, the developed PLS models for arsenic and lead quantification were utilized to evaluate these metals in the viticultural soils of Itata Valley.

3.6. Ecological Assessment of Viticultural Soils

Table 10. Statistical summary of As and Pb concentrations (mg kg⁻¹) in soils from Itata Valley.

	Median	Min	Percentile					Max	Background Upper Limit	Outliers (%)
			5	25	50	75	95
As	2.34	0.56	0.92	1.24	2.34	3.45	4.72	5.01	0.83	95.8
Pb	22.5	6.02	7.45	13.0	22.5	31.1	42.7	45.5	6.90	95.6

provides a statistical overview of the contents of As and Pb in topsoil samples. The difference between the highest and lowest concentrations of As and Pb indicates spatial diversity within the topsoil samples. This variance suggests natural disparities in soil composition among the locations sampled, potentially influenced by variations in mineral content, organic matter, geological origins, or human activities [2].

After establishing the upper limits based on the background levels, an evaluation was conducted to determine how many samples exceed these limits for both arsenic and lead, as detailed in Table 10. The results indicate that a majority of samples—approximately 95% for both elements—have concentrations exceeding these established upper limits.

Figure 6 graphically illustrates the results of the GEOI analysis, suggesting that the soils exhibit relatively low to moderate levels of contamination. This evaluation likely considers specific element concentrations in relation to their natural or background levels. EF assessments indicate a significant number of sampling points showing moderate to substantial enrichment, particularly when compared to iron concentrations, which show minimal vertical variation in the soil. The more rigorous CF index, not considering natural variations, indicates that most sampling points demonstrate moderate to considerable contamination levels for both elements. These conclusions align with the areas identified by the GEOI as moderately contaminated and reinforce their importance in terms of contamination severity.

4. Discussion

While lower RMSE values are preferable, the average percentage errors of 15.6% for arsenic and 18.9% for lead quantification using classical regression and deconvolution are considered reasonably acceptable, particularly given the analytical technique and concentration range involved. Such error magnitudes are common in this field, as evident from various studies on solid matrices [2,44,45,46].

Conversely, the utilization of the proposed PLS machine learning algorithm demonstrates improved performance, with the average percentage error notably reduced from 15.6 to 9.4% for arsenic (from an RMSEP of 5.6 mg kg⁻¹ to 3.3 mg kg⁻¹) and from 18.9% to 6.8% for lead (from an RMSEP of 12.3 mg kg⁻¹ to 5.03 mg kg⁻¹) compared to the previous univariable model. This enhanced predictive accuracy within the sample set’s concentration range is attributable to the efficiency of the multivariate calibration’s first-order advantage in quantifying in the presence of interferents, which enables the identification and use of lower-intensity signals to mitigate the errors generated by signal overlaps [46]. The RMSEP results undoubtedly demonstrate improvement; however, they do not surpass the performance achieved for liquid samples in TXRF using PLS models. Notably, the multivariate model developed by Nagata et al. (2006) [12] facilitated the determination of lead and arsenic with RMSEPs of 0.03 and 0.24 mg l⁻¹, respectively. However, the Slurry-TXRF with PLS improvements are limited in comparison to liquid sample analysis, primarily due to two factors. Firstly, liquid samples present fewer physical interferences for total reflection [10]. Achieving a comparable state from solid samples would necessitate significant resources, including time and reagents for digestion. Secondly, the TXRF equipment utilized in the aforementioned research employs synchrotron radiation TXRF (SR-TXRF), which enhances the instrument performance [12].

According to the EJCR criteria, the deconvolution method provided by the univariate calibration software for mitigating the overlap of fluorescent arsenic and lead signals appears inadequate in generating reliable quantifications, in contrast with the superior performance of the PLS method. However, as mentioned earlier, the quality of results obtained through classical calibration is commonly encountered within this analytical technique and concentration range.

Alongside the improved quality of results observed with PLS quantification, all figures of merit are also enhanced, potentially due to the multivariate calibration’s ability to exploit modeled interferents, thereby improving selectivity and reducing background noise [36].

The proposed methodology also reduces costs and analysis time compared to traditional methods [11]. This novel approach, based on Slurry-TXRF and PLS quantification models, facilitates soil monitoring for the measurement of arsenic and lead contents, and it is presented as an alternative with a positive technical and environmental impact. This is a preliminary advancement in an era where automation and unmanned systems, known for their reliability and cost-effectiveness, are becoming increasingly popular. [47].

The enhanced methodology for quantifying arsenic and lead indicates that approximately 95% of the agricultural soil samples from the Itata Valley exceed the background limit, which is expected for a soil type influenced by productive anthropogenic activity where the addition of these elements is possible through the use of fertilizers [48] and fossil fuels from machinery [49]. Additionally, the geomorphological fluvial–glacial–volcanic plain zone that the Itata Valley covers [24] could significantly contribute to the presence of these two elements in the environment, especially arsenic [50].

In general, the ecological indices suggest that the agricultural soils in the studied area are moderately contaminated by these elements, consistent with previous findings on other potentially toxic substances. However, this level of contamination does not necessarily imply immediate adverse biological effects on the resident communities or pose significant health risks to humans.

5. Conclusions

The methodology proposed in our study, which integrates Slurry-TXRF with a Partial Least Squares (PLS) machine learning algorithm, offers a significant advancement in the field of environmental monitoring, particularly for arsenic and lead contamination in soils. While the accuracy of our results does not surpass that of traditional methods, such as hydride generation for arsenic (which typically achieves errors of less than 5% with quantification limits of 0.2 mg/kg) [51] and flame atomic absorption for lead (with expected errors around 2%) [52], the application of machine learning techniques has notably improved the performance of Slurry-TXRF. Specifically, our study achieved a reduction in the average percentage error from 15.6% to 9.4% for arsenic (an RMSEP from 5.6 mg kg⁻¹ to 3.3 mg kg⁻¹) and from 18.9% to 6.8% for lead (an RMSEP from 12.3 mg kg⁻¹ to 5.03 mg kg⁻¹) just through using multivariate calibration.

These improvements, while not surpassing the precision and accuracy of the most established methods, provide significant practical advantages for environmental monitoring. The proposed methodology is faster, simpler, less costly, and more environmentally friendly, making it a valuable alternative in situations where these factors are critical.

Additionally, the PLS calibration effectively corrects the mutual influence of overlapping signals between arsenic and lead, enhancing the reliability of quantitative analyses. This is particularly advantageous in complex soil matrices where traditional methods may struggle with signal interference.

Overall, while the traditional methods remain superior in numerical accuracy, the combined use of Slurry-TXRF and machine learning represents a promising alternative that balances reasonable accuracy with substantial operational benefits. This approach is well suited for efficient environmental and ecological evaluations, particularly in regions where rapid and cost-effective monitoring is necessary.

Our assessment indicates moderate to significant contamination levels with arsenic and lead in the soils of the Itata Valley, which is typical for agriculturally used soils. However, further studies would be required to assess the potential risks to human health, including an evaluation of factors such as elemental speciation and migration from soil to plant parts, which are outside the scope of the present study.