Predicting Soil Salinity Based on Soil/Water Extracts in a Semi-Arid Region of Morocco

Abstract

Soil salinity is a major constraint to soil health and crop productivity, especially in arid and semi-arid regions. The most accurate measurement of soil salinity is considered to be the electrical conductivity of saturated soil extracts (EC_e). Because this method is labor-intensive, it is unsuitable for routine analysis in large soil sampling campaigns. This study aimed to identify the best models to estimate soil salinity based on EC_e in relation to a rapid electrical conductivity (EC) measurement in soil/water (referred to as S:W henceforward) extracts. We evaluated the relationship between EC_e and the EC_S:W extract ratios (1:1, 1:2, and 1:5) in salt-affected soils from the semi-arid Sehb El Masjoune region of Morocco. The soil salinity in this region is 0.5 to 235 dS/m, as determined by the EC_e method. A total of 125 soil samples, from topsoil (0–15 cm) and subsoil (15–30 cm) with mainly fine to medium textures, were analyzed using linear, logarithmic, and second-order polynomial regression models. The models included all samples or grouped samples according to soil texture (fine, medium) or specific textural classes. The mean EC_e values were 2.6, 3.1, and 7.9 times greater than the EC of 1:1, 1:2, and 1:5 S:W extracts, respectively. Polynomial regression models had the best predictive accuracy, R² = 0.98, and the lowest root mean square error of 10.6 to 10.7 dS/m for the EC_S:W extract ratios of 1:5 and 1:2. The polynomial models could represent the non-linear relationships between EC_e and salinity indicators, especially in the 80–170 dS/m salinity range, where other models typically underestimate the salinity. These results confirm that advanced regression techniques are suitable for predicting soil salinity in a salt-affected semi-arid region. The site-specific models outperformed previously published models, because they consider the spatial variability and heterogeneity of the salinity in the study area explicitly. This confirms the importance of calibrating soil salinity models according to the local soil and environmental conditions. Consequently, we can undertake soil salinity assessments in hundreds of samples by using the simple, rapid EC_S:W extraction method as a direct indicator of EC and extrapolate to EC_e with a polynomial regression model. Our approach enables the widespread soil salinity assessments that are needed for land-use planning, irrigation management, and crop selection in salt-affected landscapes.

1. Introduction

Approximately 15% of the world’s land has been degraded, with soil salinity accounting for 7%, predominantly in arid and semi-arid regions [1,2,3]. Salt-affected soils significantly reduce agricultural productivity [4,5], emphasizing the critical need for sustainable management strategies to mitigate this challenge.

Soil salinity increases the osmotic potential of a soil solution [6], which limits the water uptake by crops and soil microorganisms [7,8,9], damaging the health, structure, and stability of soil [10] and negatively affecting water infiltration and drainage [11]. Consequently, saline soils have lower crop yields [10]. Morocco is an arid to semi-arid country [12], where salinity constrains crop growth, with yield reductions of 19 to 26% being reported for alfalfa and similar statistics for other crops such as wheat [13,14,15]. Salinity is considered the second-most limiting factor for crop production, after erosion, in Morocco [16]. Developing and managing sustainable soil and water resources is a pressing issue affecting most countries worldwide, particularly those in arid and semi-arid regions [17]. Salinity is a common challenge in Morocco, driven by its climate of low rainfall and high evaporation rates [18], as well as management-induced factors such as the excessive use of fertilizers and saline water for irrigation [11], exemplifying the severity of this global problem.

Land managers must understand the spatial variability in soil salinity to select tolerant crops, apply soil amendments, and optimize irrigation protocols. Ideally, this is achieved according to precision agriculture principles to differentiate saline zones from areas that are not yet degraded but are susceptible to becoming saline. This requires systematic soil analysis to map the soil salinity, which is heterogeneous across fields and landscapes. Soil’s electrical conductivity (EC) is a proxy for soil salinity, because it reflects the concentration of dissolved ions in the soil. Practically, soil’s EC is usually measured in S:W extracts with ratios of 1:1, 1:2, 1:2.5, and 1:5 [19] in a cost-effective and rapid (i.e., 30 min per sample) procedure. However, S:W extracts underestimate the EC relative to the saturated paste extract method, which gives a standard EC_e value that is considered to accurately measure soil salinity [20,21,22]. Because the saturated paste extract method is time-consuming (i.e., 24 h per sample), it is costly and unsuitable for high-throughput analysis.

In theory, the EC from S:W extracts is related to the EC_e of saturated paste extracts [23]. This can be validated by comparing the EC_e and EC of S:W extracts with a specific ratio for selected samples and then applying a correction factor to adjust the S:W extracts to the standard EC_e value on an equivalent 1:1 basis. Correlations, correction factors, and equations of the relationship between EC_e and EC in S:W extracts were reported in many regions [20,21,24,25,26,27]. However, there is a disparity in these models due to site-specific variations in soil texture, the soil’s chemical composition [28], and the procedure for assessing EC in S:W extracts, such as the shaking and equilibration time [10,20,27].

Our objective was to establish a robust method to assess soil salinity from the EC of S:W extracts, which were adjusted to the standard EC_e value for the same soil. Several S:W extraction ratios were tested (i.e., 1:5, 1:2, and 1:1) in topsoil and subsoil across a topographically complex area spanning hundreds of km² in semi-arid Morocco. We also discuss the factors affecting the site-specific salinity model in this area.

2. Materials and Methods

2.1. Study Area

The study area is located in an agricultural area around Sehb ElMasjoune (centered at 43°38′25.74″ N; 80°24′36.64″ W; Figure 1), about 60 km north of Marrakech, Morocco. The investigated site covers an area of about 98,100 hectares (ha). The Sehb ElMasjoune area is in a topographic depression with a central dry lake (3800 ha), surrounded by the highlands of Gantour from the north and Jebilat Mountains from the south [29]. The topography is relatively flat across the Sehb ElMasjoune area, with slopes from 0.1% to 3% and an elevation between 401 and 500 m above sea level.

Heavy rains can periodically transform the central lake into an occasional wetland, recharged with surface runoff from the surrounding landscape. Usually, the accumulated water in the lake evaporates quickly (i.e., in a few weeks), which leaves salt deposits that salinize the soil within the topographic depression and the surrounding agricultural lands [29].

The Sehb ElMasjoune region is characterized by the presence of various natural halophyte species, including Aizoanthemopsis hispanica (L.) Klak, Frankenia pulverulenta L., Salsola soda L., Haloxylon scoparium Pomel, Atriplex semibaccata R.Br., Herniaria hirsuta L., Limonium lobatum (L.f.) Chaz., Spergularia bocconei (Scheele) Graebn., Peganum harmala L., and Anabasis articulata (Forsskal) [30], which are distributed sporadically across the sparsely vegetated landscape. Similar to other nearby areas with the same climate, the agriculture in this region primarily consists of traditional farmlands, where cereals such as barley and wheat, as well as olive trees, are cultivated [31]. However, the northern part of the lake shows a higher soil salinity compared to the southern areas. It is worth noting that a significant portion of agricultural land in the region has been abandoned, likely due to the combined effects of drought and soil salinization. These abandoned lands are gradually being naturally recolonized by the same halophyte species mentioned above (Figure 2).

The Sehb El Masjoune region experiences a semi-arid climate, characterized by short, mild winters (November to April) and long, hot summers. The average daily temperature ranges from 19 °C in January to 37 °C in July, with extremes ranging from a minimum of approximately 0 °C to a maximum of 46 °C. The annual precipitation is less than 200 mm, while the region records an average relative humidity of around 53%, an annual evaporation rate of 2253 mm, and an evapotranspiration rate of approximately 2700 mm [32]. The soil texture in the topsoil is fine (58%) to medium (40%), with few (2%) coarsely textured soils, a pH from 7.6 to 9.1, and EC_e values from 0.52 to 235 dS/m. Further details about the study area are reported elsewhere [16,29].

2.2. Soil Sample Collection and Electrical Conductivity Analyses

In June 2023, we collected 125 soil samples from the Sehb ElMasjoune area (Figure 1) along a north–south gradient to account for the spatial variability in soil salinity. At each location, we established a 1 m² quadrat; then, five subsamples (i.e., 1.5 kg of soil each) were collected per quadrat from the topsoil (0–15 cm) using hand trowels. One composite topsoil sample was made by mixing the 5 subsamples (i.e., 4 corners + center) within the 1 m² quadrat to obtain a representative topsoil sample, based on the observed variation in salinity at the soil surface. In the same 1 m² quadrat, the subsoil sample (15–30 cm) was 1.5 kg of soil taken from the center of the quadrat using a shovel. This is justified by (1) the relatively stable and homogenous salinity in the subsoil (15–30 cm) compared to the topsoil (0–15 cm) and (2) the fact that the geolocated subsoil sampling point was used to correlate the field-measured soil salinity (EC_e) with the proximal sensing data (i.e., collected with the EM38 electromagnetic induction method) in a companion study. All soil samples were oven-dried (38 °C for 24 h). Then, crop roots and residues were removed, the soil was ground < 200 µm, and the sample underwent chemical and physical analyses. The soil texture (i.e., sand, silt, and clay percentages) was measured by the hydrometer method, and soil textural classes were assigned using the USDA soil textural triangle. S:W extracts were prepared by adding distilled water to 20 g soil in a plastic container at the following ratios: 20 mL for the EC_1:1 (n = 101 samples), 40 mL for EC_1:2 (n = 101 samples), and 100 mL for EC_1:5 (n = 125 samples). Then, the containers were shaken mechanically for 30 min using a horizontally oscillating platform shaker and filtered through a Whatman 2 qualitative cellulose fiber filter. The filtered extracts were analyzed within 3 min on a calibrated electrical conductivity meter (pH/Cond meter SevenDirect SD23, Mettler-Toledo (S) Pte Ltd., Greifensee, Switzerland), and all the EC values were expressed in dS/m.

To prepare a saturated paste for each sample, the soil was first oven-dried and sieved to remove large particles. Subsequently, 200 g of the prepared soil was incrementally mixed with distilled water with continuous stirring until the paste reached the saturation point, characterized by a glistening surface without free-flowing water. This procedure followed the protocol of Rhoades [33]. The saturated soil paste was covered and equilibrated for 24 h. Following equilibration, the paste was centrifuged for 10 min to separate the liquid phase. The liquid extract was subsequently collected using a vacuum pump, and its electrical conductivity (EC_e, dS/m) was determined on a calibrated electrical conductivity meter.

2.3. Statistical Analysis of Data

The EC rates of the three S:W extract ratios were compared using a t-test, with statistical significance determined at p < 0.05. To analyze the relationship between an S:W extract’s EC and EC_e, regression analysis was conducted. The model was calibrated using 75% of the data, and three models were tested—linear regression, linear regression of log10-transformed data, and second-order polynomial regression—and referred to as Equations (1)–(3), respectively. The remaining 25% of the dataset was reserved for model validation and performance assessment. Since the soil texture can impact the EC_S:W [34], the relationships between EC_e and EC_S:W were tested with texture-specific data groupings. The first group was composed of all samples with a full range of soil textures (named AllS; Figure 3). The second grouping was based on two soil texture groups (STGs) consisting of finely textured soils (clay, clay loam, sandy clay loam, silty clay, and silty clay loam) and medium-textured soil (loam, silt loam). The third grouping used soil texture classes (STs), which included clay, clay loam, and silt loam. The STG was grouped according to the USDA soil texture classification [35]. Regression models were validated using 25 samples for EC_1:5 in the AllS split and 22 samples for both EC_1:2 and EC_1:1 in the AllS split. For the STG and ST splits, model validation for EC_1:1 and EC_1:2 was performed using 21 and 18 samples, respectively, and 24 and 20 samples, respectively, for EC_1:5 (

Table 1. The number of samples used for model validation according to the soil-to-water (S:W) ratios in the EC extracts.

Split	S:W Ratio	No of Validation Samples
AllS	1:1	22
	1:2	22
	1:5	25
STG	1:1	21
	1:2	21
	1:5	24
ST	1:1	18
	1:2	18
	1:5	20

). The number of validation samples in each grouping varied according to the number of specific soil textures that were available in the dataset.

We confirmed that the samples used in the validation process were not included in calibrating the models. The EC distributions of the calibration and validation sets of the three EC_S:W, presented in Figure 3, show that the EC_e values of the validation dataset are representative of the range in the calibration dataset. Model statistics include the coefficient of determination (R²; Equation (4)), the root mean square error (RMSE; Equation (5)), and the Nash–Sutcliffe prediction efficiency (NSE; Equation (6)). These test statistics reflect the goodness-of-fit in each regression model.

$${\mathrm{E}\mathrm{C}}_{\mathrm{e}}=\mathrm{a}\times {\mathrm{E}\mathrm{C}}_{\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}:\mathrm{w}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}}\pm \mathrm{b}$$

(1)

$${\mathrm{E}\mathrm{C}}_{\mathrm{e}}={10}^{(\mathrm{a}\times {\mathrm{L}\mathrm{o}\mathrm{g}}_{10}({\mathrm{E}\mathrm{C}}_{\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}:\mathrm{w}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}})\pm \mathrm{b}}$$

(2)

$${\mathrm{E}\mathrm{C}}_{\mathrm{e}}=\mathrm{c}\times {{(\mathrm{E}\mathrm{C}}_{\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}:\mathrm{w}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}})}^2\pm \mathrm{d}\times {\mathrm{E}\mathrm{C}}_{\mathrm{s}\mathrm{o}\mathrm{i}\mathrm{l}:\mathrm{w}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}}\pm \mathrm{e}$$

(3)

$${\mathrm{R}}^2=\left({\displaystyle \frac{{\sum }_{\mathrm{i}=1}^n\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}{\sqrt{{{\sum }_{\mathrm{i}=1}^n\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)}^2+{\sum }_{\mathrm{i}=1}^n{\left({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}}\right)}^2}}}\right)$$

(4)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^n{\left({\mathrm{x}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}{n}}}$$

(5)

$$\mathrm{N}\mathrm{S}\mathrm{E}=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^n{\left({\mathrm{x}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}{{\sum }_{\mathrm{i}=1}^n{\left({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}\right)}^2}}$$

(6)

where

x_i and y_i are the measured and estimated values, respectively.

$\overline{\mathrm{x}}$ and $\overline{\mathrm{y}}$ represent the means of the measured and predicted values, respectively.

n is the number of samples.

a is the equation slope.

b is the intercept.

c, d, and e are the polynomial second-order factors.

3. Results and Discussions

3.1. Soil Texture Salinity Characterizations

Table 2. Properties of the soil samples gathered from the Sehb ElMasjoune area.

Soil Texture and Soil Texture Groups	ECe		EC1:1		EC1:2		EC1:5
	NS *	% of NS	NS	% of NS	NS	% of NS	NS	% of NS
Clay (C)	28	22	20	20	20	20	28	22
Clay Loam (CL)	40	32	34	33	34	33	40	32
Loam (L)	27	22	21	21	21	21	27	22
Sandy Clay Loam (SCL)	4	3	2	2	2	2	4	3
Sandy Loam (SL)	3	2	3	3	3	3	3	2
Silt Loam (ZL)	19	15	17	17	17	17	19	15
Silty Clay Loam (ZCL)	4	4	4	4	4	4	4	4
Finely Textured Soil (FTS)	72	58	58	57	58	57	72	58
Medium-Textured Soil (MTS)	50	40	40	40	40	40	50	40
Coarsely Textured Soil (CTS)	3	2	3	3	3	3	3	2

* Number of samples.

shows the main textures of all the samples based on the USDA soil texture classification. The analyzed soils are clay loam (32%), clay (22%), and loam (22%), followed by silt loam (15%), sandy clay loam (3%), silty clay loam (3%), and sandy loam (2%). Therefore, fine-, medium-, and coarse-textured soils represent 58%, 40%, and 2% of the studied samples, respectively.

Table 3. Statistic description of the salinities of the soil samples used in this study.

	* ECe (dS/m)	* EC1:1 (dS/m)	* EC1:2 (dS/m)	* EC1:5 (dS/m)
Count	125	101	101	125
Mean	87	34	28	11
Std	70	30	27	12
Min	0.5	0.23	0.14	0.07
Q1	10	13	8	2
Median	79	29	22	10
Q3	131	44	33	13
Max	235	131	100	47

* EC_e: electrical conductivity of the extracted saturated paste; EC_1:1;1:2;1:5: electrical conductivity of the S:W ratio.

shows descriptive statistics of the EC_e and EC_S:W of the studied samples. The mean values of the EC_e, EC_1:1, EC_1:2, and EC_1:5 were found to be 87, 34, 28, and 11 dS/m, respectively. Only the means of EC_1:1 and EC_1:2 were not significantly different at p < 0.05. With 74%, the extremely saline soil (EC_e ≥ 16 dS/m) represented most of the sampling locations in the investigated area. Additionally, 6% were saline soils (4 dS/m < EC_e ≤ 16 dS/m), and the remaining 20% were non-saline to slightly saline soils (EC_e ≤ 4 dS/m).

3.2. Relationship Between the EC_e and the EC in S:W Extracts

The mean EC_e was 7.9-fold greater than the mean EC_1:5, 3.1 times greater than the EC_1:2, and 2.6-fold larger than the EC_1:1 in the study area (Table 3). This is consistent with reports that found that S:W extracts have lower EC values than the EC_e value of the same soil [10,26,36]. The 3.1-fold difference between EC_e and EC_1:2 is approximately the same as that reported by Spiteri et al. [11]. When the EC_e was <170 dS/m, there was a linear relationship between the EC_e and the EC values of the tested S:W extracts, but after that, there was a curvilinear relationship (Figure 4). This was interpreted as a dilution process, where the EC value declines with an increasing proportion of water in the S:W extract, as noted in the literature [19,20].

The EC_e had a linear relationship with the EC of the S:W extracts, and the R² values were from 0.78 to 0.92 for EC_1:5, 0.77 to 0.95 for EC_1:2, and 0.79 to 0.95 for EC_1:1 (

Table 4. Regression equations between EC_e and EC_S:W extracts used in this study, along with the R² of each one.

		AllS 1	STG 2		ST 3
			FTS 4	MTS 5	Clay	Clay Loam	Loam	Silty Loam
Linear (L)	EC1:5	y = 5.2884 × EC1:5 + 24.891 R2 = 0.85	y = 5.3486 × EC1:5 + 29.076 R2 = 0.78	y = 5.2673 × EC1:5 + 18.513 R2 = 0.90	y = 5.1491 × EC1:5 + 21.939 R2 = 0.82	y = 8.7411 × EC1:5 + 7.5814 R2 = 0.92	y = 9.2169 × EC1:5 + 2.7519 R2 = 0.90	y = 4.5641 × EC1:5 + 38.128 R2 = 0.88
	EC1:2	y = 2.4634 × EC1:2 + 23.927 R2 = 0.88	y = 2.5051 × EC1:2 + 29.348 R2 = 0.77	y = 2.4543 × EC1:2 + 14.562 R2 = 0.95	y = 2.5059 × EC1:2 + 18.827 R2 = 0.81	y = 3.7228 × EC1:2 + 12.16 R2 = 0.85	y = 4.1179 × EC1:2 + 2.2914 R2 = 0.93	y = 2.0411 × EC1:2 + 45.074 R2 = 0.93
	EC1:1	y = 2.2602 × EC1:1 + 17.012 R2 = 0.90	y = 2.4286 × EC1:1 + 17.279 R2 = 0.83	y = 2.2009 × EC1:1 + 10.576 R2 = 0.95	y = 2.3942 × EC1:1 + 10.899 R2 = 0.87	y = 2.8176 × EC1:1 + 12.006 R2 = 0.79	y = 3.3068 × EC1:1 + 1.0312 R2 = 0.95	y = 1.8792 × EC1:1 + 36.906 R2 = 0.88
Log-Transformation (LT)	EC1:5	y = 0.9439 × Log(EC1:5) + 0.960 R2 = 0.95	y = 0.9104 × Log(EC1:5) + 1.0154 R2 = 0.96	y = 0.9525 × Log(EC1:5) + 0.9116 R2 = 0.95	y = 0.8606 × Log(EC1:5) + 1.0105 R2 = 0.95	y = 0.9521 × Log(EC1:5) + 1.0296 R2 = 0.96	y = 1.0142 × Log(EC1:5) + 0.9572 R2 = 0.93	y = 1.1379 × Log(EC1:5) + 0.59 R2 = 0.88
	EC1:2	y = 0.9619 × Log(EC1:2) + 0.6113 R2 = 0.98	y = 0.9718 × Log(EC1:2) + 0.609 R2 = 0.96	y = 0.9504 × Log(EC1:2) + 0.6034 R2 = 0.99	y = 0.8938 × Log(EC1:2) + 0.6514 R2 = 0.75	y = 1.0269 × Log(EC1:2) + 0.5996 R2 = 0.98	y = 1.0674 × Log(EC1:2) + 0.6342 R2 = 0.99	y = 0.8053 × Log(EC1:2) + 0.8085 R2 = 0.93
	EC1:1	y = 1.0118 × Log(EC1:1) + 0.441 R2 = 0.99	y = 1.0272 × Log(EC1:1) + 0.4327 R2 = 0.97	y = 0.9955 × Log(EC1:1) + 0.4346 R2 = 0.99	y = 0.981 × Log(EC1:1) + 0.4615 R2 = 0.86	y = 1.0558 × Log(EC1:1) + 0.4274 R2 = 0.98	y = 1.09 × Log(EC1:1) + 0.4484 R2 = 0.99	y = 0.9353 × Log(EC1:1) + 0.5066 R2 = 0.95
Polynomial (P)	EC1:5	y = −0.1227 × (EC1:5)2 + 10.36 × EC1:5 + 0.9916 R2 = 0.94	y = −0.1278 × (EC1:5)2 + 10.246 × EC1:5 + 1.5043 R2 = 0.89	y = −0.1348 × (EC1:5)2 + 11.014 × EC1:5 + 0.5157 R2 = 0.97	y = −0.0736 × (EC1:5)2 + 8.052 × EC1:5 + 4.331 R2 = 0.85	y = −0.2874 × (EC1:5)2 + 13.36 × EC1:5 − 1.4652 R2 = 0.95	y = −0.6149 × (EC1:5)2 + 17.427 × EC1:5 − 2.8076 R2 = 0.97	y = −0.1523 × (EC1:5)2 + 12.34 × EC1:5 − 23.396 R2 = 0.98
	EC1:2	y = −0.023 × (EC1:2)2 + 4.5092 × EC1:2 + 1.6773 R2 = 0.94	y = −0.0257 × (EC1:2)2 + 4.6139 × EC1:2 + 0.1577 R2 = 0.84	y = −0.0257 × (EC1:2)2 + 4.7934 × EC1:2 + 1.6606 R2 = 0.99	y = −0.0369 × (EC1:2)2 + 6.0741 × EC1:2 − 43.427 R2 = 0.83	y = −0.0761 × (EC1:2)2 + 6.6502 × EC1:2 − 4.677 R2 = 0.91	y = −0.1299 × (EC1:2)2 + 7.7707 × EC1:2 − 1.7488 R2 = 0.99	y = −0.0249 × (EC1:2)2 + 4.7535 × EC1:2 − 1.0267 R2 = 0.99
	EC1:1	y = −0.0135 × (EC1:1)2 + 3.6088 × EC1:1 − 0.9008 R2 = 0.94	y = −0.014 × (EC1:1)2 + 3.6093 × EC1:1 − 0.9266 R2 = 0.85	y = −0.0145 × (EC1:1)2 + 3.7243 × EC1:1 − 1.0792 R2 = 0.99	y = −0.0281 × (EC1:1)2 + 5.3768 × EC1:1 − 48.62 R2 = 0.91	y = −0.0415 × (EC1:1)2 + 4.8592 × EC1:1 − 2.8368 R2 = 0.84	y = −0.0579 × (EC1:1)2 + 5.2633 × EC1:1 − 1.8258 R2 = 0.98	y = −0.0183 × (EC1:1)2 + 4.3581 × EC1:1 − 25.237 R2 = 0.97

¹ All samples, ² soil texture groups, ³ soil texture, ⁴ finely textured soil, ⁵ medium-textured soil.

). The association between EC_e and the log-transformed EC of the S:W extracts tended to have higher R² values, up to 0.96 (EC_1:5) and 0.99 (EC_1:2, EC_1:1; Table 4).

Due to the curvilinear relationship between the EC_e and EC in the S:W extracts (Figure 3), the polynomial curve fits well to the data, with R² values between 0.85 and 0.98 for EC_1:5, between 0.83 and 0.99 for EC_1:2, and between 0.84 and 0.99 for EC_1:1 (Table 4). This is consistent with [21,37], who also reported a curvilinear relationship between EC_e and EC_S:W extracts.

As indicated in Table 4, the slope of the developed models is higher when transitioning from medium- to fine-textured soil, which is similar to what Leksungnoen et al. [22] and Sonmez et al. [26] found. This may be explained as soils with finer textures, which are rich in clay, possess a superior capacity for water retention, which significantly influences ion retention, where a heightened clay proportion correlates with diminished ion release and lower soluble salt concentrations in the soil solution, manifesting a dilution effect. Conversely, salts are more readily leached into soils with medium to coarse textures, resulting in elevated soluble salt concentrations in the soil solution [22,26,38]. Nonetheless, this established pattern was not observed in silt loam, which may be attributed to its limited sample size or the impact of an additional factor affecting the water and solute retention properties in these soil types. These possibilities remain to be investigated.

3.3. Soil EC Models Compared to Those Used for the Sehb ElMasjoune Area of Morocco

The correlation between the EC_e in the saturated soil paste extracts and EC_S:W was reported in previous studies and is summarized in

Table 5. Correlations between soil-saturated paste extracts’ electrical conductivity (EC_e) and various soil-to-water extracts’ electrical conductivities (EC_S:W), as reported by different researchers, in conjunction with the respective ranges of EC_e values and the scope of the study.

Reference	Regression Equation	EC Range (dS/m)	Country/Region
Herrero and Pérez-Coveta, 2005 [18]	ECe = 7.63 × EC1:5 − 0.51	2.9–4.6	Spain/Ebro basin
Ozcan et al., 2006 [39]	ECe = 1.93 × EC1:1 − 0.57 ECe = 5.97 × EC1:5 − 1.17	-	Turkey
Sonmez et al. 2008 [26]	ECe = 8.22 × EC1:5 − 0.33 ECe = 7.58 × EC1:5 + 0.06 ECe = 7.36 × EC1:5 − 0.24	0.22–17.68	Turkey/Antalya
	ECe = 3.68 × EC1:2.5 + 0.22 ECe = 3.84 × EC1:2.5 + 0.35 ECe = 4.34 × EC1:2.5 + 0.17
	ECe = 2.03 × EC1:1 − 0.41 ECe = 2.15 × EC1:1 − 0.44 ECe = 2.72 × EC1:1 − 1.27
Visconti and Miguel De Paz, 2012 [40]	ECe = 5.7 × EC1:5 − 0.2	0.3–3.3	Spain/Southeast
Khorsandi and Yazdi 2011 [25]	ECe = 5.37 × EC1:5 + 0.57 ECe = 5.60 × EC1:5 − 4.37	0.48–171.3	Iran/Yazd
He et al., 2013 [37]	ECe = 2.86 × EC1:5 + 2.96	0.0–17.0	USA/North Dakota
Klaustermeier et al., 2016 [21]	LogECe = 1.2562 × Log(EC1:5) + 0.7659	0.4–126.0	USA/North Dakota
Aboukila and Norton, 2017 [24]	ECe = 5.04 × EC1:5 + 0.37	0.62–10.3	Egypt/Beheira
Aboukila and Abdelaty 2017 [36]	ECe = 7.46 × EC1:5 + 0.43	0.29–18.35	Egypt/Beheira
Leksungnoen et al., 2018 [22]	ECe = 5.99 × EC1:5 + 0.62	12.4–80.7	Thailand/Khorat and Sakhon basins
Kargas et al., 2018 [19]	ECe = 6.53 × EC1:5 − 0.108 ECe = 1.83 × EC1:1 + 0.117	0.47–37.5	Greece/multiple locations
Kargas et al., 2020 [41]	ECe = 6.58 × EC1:5	0.61–25.9	Greece/three locations
USDA (1954) [35]	ECe = 3 × EC1:1	-	USA
Hogg and Henry 1984 [20]	ECe = 1.56 × EC1:1 − 0.06	0.10–42.01	Canada/Saskatoon
Spiteri and Sacco in 2024 [11]	ECe = 100.972×Log (EC1:2)+0.515 ECe = 101.048×Log (EC1:2)+0.456	0.73–26.73	Islands of Malta
Zhang et al. 2005 [27]	ECe = 1.79 × EC1:1 + 1.46	0.165–108	USA/Oklahoma and Texas
Chi and Wang, 2010 [42]	ECe = 11.68 × EC1:5 − 5.77 ECe = 11.04 × EC1:5 − 2.41 ECe = 11.74 × EC1:5 − 6.15	1–227.0	China/Songnen
Park et al. 2019 [43]	ECe = 8.70 × EC1:5	-	Republic of Korea/south-western

along with references, soil salinity range, and the region or country in which those studies were conducted. Most of these studies have reported a strong linear correlation between different EC_S:W and EC_e values. Few of the studies in Table 5 showed a non-linear relationship [18,21,37].

3.4. Model Performance Comparison and Assessment

To investigate the relationships between the EC_e and EC values from soil/water extracts in arid and semi-arid regions, various regression models were considered, and the goodness-of-fit was evaluated statistically. To validate these models, a performance evaluation was conducted using an independent dataset, as detailed in Section 2.3, which was not included in the initial calibration phase (Figure 3).

3.4.1. Linear Models

The linear regression models demonstrated consistent performances across the EC_1:1, EC_1:2, and EC_1:5 ratios, showing high R² values (indicative of a strong linear relationship between the actual and predicted EC_e values). Despite this, discrepancies were observed in other metrics, particularly the NSE and RMSE, which displayed significant variability across dataset splits. This variation suggests that linear models, while capable of capturing overall trends, may struggle to account for the complexity of the relationships between soil salinity variables. For example, although certain linear models provided high predictive accuracy, their robustness diminished when the salinity gradients were complex or highly variable. These observations align with findings by He et al. [37], underscoring the importance of supplementing linear analysis with non-linear or transformed models for more reliable predictions.

3.4.2. Logarithmic Transformation Models

Logarithmic transformation enhances a models’ ability to capture non-linear relationships, as evidenced by the high R² values (up to 0.99) (Figure 5). This highlights the effectiveness of log-transformed models in handling datasets with significant variability. However, the performance of these models is not uniform across all conditions, as they may face challenges in providing consistent accuracy due to the compression of the higher values that are inherent in the transformation process. These findings suggest that while log-transformed models are valuable for addressing non-linearities and variability, their application must be carefully considered to ensure reliable performances across diverse datasets.

3.4.3. Polynomial Models

The polynomial models exhibited remarkable performances across the three S:W extract ratios (EC_1:1, EC_1:2, and EC_1:5). These models consistently achieved high R² values (above 0.96), denoting strong predictive capabilities, alongside high NSE values and low RMSE values (below 15.75 dS/m). This robust performance highlights the ability of polynomial regression to accurately capture complex patterns in soil salinity datasets. Among the different extracts, the polynomial models based on the EC_1:2 and EC_1:5 ratios exhibited outstanding performances across various splits, with particularly noteworthy results in the STG and ST splits. For the EC_1:2 ratio, both the R² and NSE metrics exceeded 0.97, while the RMSE remained below 13.41 dS/m, highlighting a precise fit with minimal errors. Similarly, the polynomial model developed for the EC_1:5 ratio under the ST split achieved exceptional performance, with R² and NSE values reaching 0.98 and a remarkably low RMSE of 10.64 dS/m. These findings underscore the efficacy and adaptability of polynomial regression in handling datasets with varying salinity gradients, making the EC_1:2 and EC_1:5 extracts particularly reliable for modeling soil salinity with high precision and minimal error. Interestingly, comparisons between original and logarithmically transformed data revealed nuanced differences in performance. While the log-transformed models often yielded higher R² values, the polynomial models based on the original data consistently outperformed them in terms of NSE and RMSE, particularly for the AllS and STG splits, highlighting their robustness and reliability for soil salinity prediction. An exception was observed with the log-transformed EC_1:2 dataset, which performed comparably to the polynomial models, achieving R², NSE, and RMSE values of 0.98, 0.97, and 11.94 dS/m, respectively. This indicates that logarithmic transformations can occasionally match the performance of polynomial models, although the latter generally provide a more balanced and consistent performance across all metrics (Figure 5).

To illustrate additional details concerning the resemblance of the developed models, the optimal linear models utilizing original and logarithmically transformed data, along with the high-performing polynomial models from the three splits, were selected and plotted alongside each other, as shown in Figure 6. This figure presents scatter plots comparing the measured and predicted EC_e values for S:W extract ratios (1:5, 1:2, and 1:1) across three modeling approaches: L, LT, and P. The models demonstrated improved predictive performance with higher S:W dilution ratios, as 1:5 and 1:2 consistently yielded higher R² values and lower RMSE values compared to the 1:1 ratio. Among the models, the P approach outperformed the L and LT approaches. This reinforces the notion that polynomial models excel in capturing the inherent non-linear relationships in the dataset, making them the most effective among the evaluated approaches.

Among the fifteen models evaluated below, the polynomial models based on the ST split across both EC_1:2 and EC_1:5 exhibited the highest and most similar assessment metrics, with an R² of 0.98 for each; the electrical conductivity readings for EC_1:2 and EC_1:5 were recorded at 10.71 dS/m and 10.64 dS/m, respectively. This indicates that the ST split effectively enhances model performance, providing high predictive reliability across different S:W ratios. There was a slight outperformance of EC_1:2 concerning the calibration dataset. On the other hand, the model derived from log-transformed EC_1:2 data in the ST split demonstrated satisfactory results in the validation part. However, relatively unsatisfactory results were observed in the calibration part. With the exception of the LT-ST_1:2, the linear models based on the original and log-transformed data produced the relatively poorest results, with an R² ranging from 0.91 to 0.95 and the RMSE falling within the range of 17.35–23.43 dS/m. This underlines the limitations of linear models for representing complex patterns of salinity, especially when compared to polynomial approaches.

Some other studies found that prediction inaccuracies may increase as the EC_e range shifts from one level to another [21,22,37]. The EC_e values were categorized into three ranges to assess the reliability of the chosen models in each range (Figure 6). Among the diverse ranges, the intermediate range, specifically between 80 and 170 dS/m, was distinguishable by producing the highest RMSE values in all the models, in contrast to the initial (0.52–80 dS/m) and last (170–235.20 dS/m) ranges, which exhibited better RMSE values. This observation aligns with findings in prior studies [21,22,37], emphasizing that the intermediate range is characterized by higher variability and complexity than at the lower and higher ends of the salinity gradient, posing a challenge for predictive accuracy. Alongside satisfactory metrics in the initial and last ranges, the superior performances of the P-ST_1:5 and P-ST_1:2 models were noted (Figure 6) in the intermediate range compared to the other models; nevertheless, a slight underestimation of EC_e was observed in this range, which is consistent with the outcomes of [21]. The P-STG_1:2, P-AllS_1:1, and P-STG_1:1 models demonstrated good performance in the last range. Regarding the linear models, the predictions showed an underestimation of EC_e values in the intermediate range. This suggests that linear models may lack the flexibility to handle the nuances of intermediate EC_e values, while polynomial models provide a more robust alternative. As polynomial models predict negative values when the EC_S:W values are exceedingly small and display an overestimation of EC_e around 80 dS/m, the LT-ST_1:2 model demonstrated the most optimal performance in the initial range. Consequently, a synergetic utilization of the two models P-ST_1:2 and LT-ST_1:2 as a hybrid model (Hybrid-ST_1:2) could improve the precision of the prediction results. This proposed hybrid approach capitalizes on the strengths of both models, offering a tailored solution to address range-specific prediction challenges.

To evaluate the effectiveness of various predictive models in determining EC_e, the scatter plot (Figure 7) illustrates a visual representation of the best models from the present research compared to superior ones established in previous studies (

Table 6. Statistical comparison of the generated models in the current study with the models developed in previous studies.

Model	Reference	Equation	R2	NSE	RMSE (dS/m)	Soil Propriety	EC Range (dS/m)
SPITERI-11:2	Spiteri and Sacco (2024) [11]	ECe = 101.048×Log (EC1:2)+0.456	0.93	0.90	24.54	Fine soil	0.73–26.73
SPITERI-21:2	Spiteri and Sacco (2024) [11]	ECe = 100.972×Log (EC1:2)+0.515	0.91	0.86	29.64	All soil
Aboukila1:5	Aboukila and Norton (2017) [24]	ECe = 5.04 × EC1:5 + 0.37	0.91	0.82	33.44	Fine soil	0.62–10.3
Kargas1:5	Kargas et al. (2018) [41]	ECe = 6.53 × EC1:5 − 0.108	0.91	0.86	29.68	Fine soil	0.47–37.5
Khorsandi1:5	Khorsandi and Yazdi (2011) [25]	ECe = 5.37 × EC1:5 − 0.57	0.91	0.86	29.94	All soil	0.48–171
Ozcan1:5	Ozcan et al. (2006) [39]	ECe = 5.97 × EC1:5 − 1.17	0.91	0.87	27.98	All soil	-
Visconti1:5	Visconti Reluy and De Paz (2012) [40]	ECe = 5.7 × EC1:5 − 1.20	0.91	0.87	28.36	All soil	0.3–3.3
P-ST1:5	This study	See Table 4	0.98	0.98	10.64	Fine and medium soil	0.52–235.20
P-ST1:2					10.71
Hybrid-ST1:2					10.08

). The validation dataset employed in the present research was used to carry out the comparisons. Among all the examined models, the investigations carried out by Spiteri et al. and Kargas et al. [11,19] exhibited superior effectiveness (Figure 7 and Table 6) when compared to the previously mentioned models in the existing literature. The models of these investigations achieved an R² varying from 0.91 and 0.93, as well as an NSE within the range of 0.82–0.90 and an RMSE fluctuating from 24.54 to 33.44 dS/m. In comparison, the polynomial models (P-ST_1:5 and P-ST_1:2) in the present study significantly outperformed these benchmarks, highlighting their potential for advancing the accuracy of soil salinity modeling.

The performances of these models were analyzed based on the overall trend. Similar to our findings, EC_1:2 and EC_1:5 extract models were found to be more appropriate in several studies, contrasting with models based on the EC_1:1 extract. The proximity of the data points to the 1:1 fit line for most models between 0 dS/m and 40 dS/m, which implies a generally high level of prediction accuracy. As shown in Table 4, the slopes of the chosen EC_1:5 extract models were between 5.04 and 6.53, which confirms the dominance of finely textured soil in the used dataset [22,38]. This result highlights the importance of soil texture in shaping the predictive capabilities of the models, emphasizing the necessity to tailor modeling approaches based on soil characteristics. Meanwhile, models such as those presented in [25,36,39,40] demonstrated a particularly low bias beyond 200 dS/m, indicating their robustness in predicting EC_e across this range. On the other hand, the All-samples model of Spiteri et al. and Kargas et al.’s [11,41] model exhibit an overestimation of the predictive values of EC_e in the last range. Despite the satisfactory performance in some areas, the models from the literature significantly underestimate EC_e values within the range of 40–170 dS/m. This discrepancy in the intermediate range underscores the limitations of existing models in accurately capturing the variability of EC_e values, particularly in regions where the relationship between predictors and EC_e becomes more complex. This disagreement can be attributed to the linear approach taken in modeling a relationship that is essentially curvilinear, displaying similar trends to those observed in our constructed linear models (Figure 6). In the initial range, our data demonstrate an overestimation originating from our linear models’ relatively high intercept values. Conversely, all linear regression models from prior research accurately predict the EC_e in the initial range, underestimate EC_e in the intermediate range, and overestimate EC_e in the last range. These trends highlight the inherent limitations of linear regression in capturing the nuanced, non-linear relationships between predictors and EC_e values across a wide salinity gradient. This disagreement is related to the observation that the linear models developed in the previous research were derived from a more restricted EC_e range than ours, corresponding to the range in which our data show a linear correlation (0–170 dS/m). The broader range of EC_e values in the current dataset presents additional challenges but also provides an opportunity to develop more generalizable models. Considering this, the log-transformed models developed by Spiteri et al. [11] based on the EC_1:2 extract of finely textured soil were similar to our findings. This is further confirmation that log-transformed models can reflect non-linear patterns in EC_e prediction, particularly for finely textured soils.

3.5. The Development of an Automation Conversion Tool of the EC Ratio Method to EC_e

To simplify the selection of a regression model for predicting EC_e, we developed a Python script that can automate the EC_e predictions. By specifying the percentage of each soil texture fraction and selecting the EC_S:W extract ratio (EC_1:2 or EC_1:5), users can predict EC_e values in a study area with an accuracy of R² = 0.98. This calculator was a logical extension of the work presented here, but it was not used to generate the results presented in this work. It should be easy to use in a scientific context to achieve better EC_e predictions in agricultural and environmental studies with soil and climatic conditions, like those in the Sehb El Masjoune region. The code is available upon request to any interested researcher or stakeholder.

4. Conclusions

This study tested regression models for their suitability for predicting the EC_e from various EC_S:W extract ratios, namely EC_1:1, EC_1:2, and EC_1:5, for the semi-arid Sehb El Masjoune region in Morocco. Among all the tested approaches, the polynomial regression models showed the best predictive performance, yielding the highest R² values of 0.98, with RMSE < 15.75 dS/m and NSE values exceeding 0.96. The EC_1:5 extract ratio most closely fit the predicted EC_e line in all models; thus, it is recommended for rapid soil salinity assessments in this area.

This study demonstrates that advanced regression techniques are suitable for optimizing the EC_S:W extract ratios, because they can represent the non-linear relationships between the EC_e and salinity indicators effectively. Polynomial models are suitable for representing the complex, curvilinear relationships between soil salinity and EC_S:W extract ratio values, which often cannot be modeled with a linear model or a logarithmic transformation. This appears to be of particular importance in landscapes with variable soil textures (e.g., from fine to medium, with three dominant soil classes) and a wide salinity gradient of three orders of magnitude from 0.5 to 235 dS/m. We recommend using polynomial regression to describe the EC variation for better estimates and more accurate predictions in soil salinity assessments across arid and semi-arid regions.

This study is important for agricultural and environmental applications in semi-arid regions. The accurate prediction of EC_e supports improved land-use planning, optimization of irrigation practices, and appropriate crop selection in areas that are vulnerable to productivity loss due to salinity. The Python-based script that was developed from the best-fit polynomial models can be adopted as a decision-support tool for researchers and stakeholders who require information on soil salinity. Still, the equations developed in this study must be validated for other locations. Future research should be directed at expanding the datasets into different types of soils and environmental conditions. We are optimistic that machine learning methods could further improve the predictive capability of soil salinity from direct measurements and proximal analysis for more robust soil salinity assessment in the future. Overall, this work provides a significant advance in soil salinity prediction, offering practical tools and insights that bridge the gap between research and application in arid and semi-arid regions.