Predicting Soil Organic Carbon from Sentinel-2 Imagery and Regional Calibration Approach in Salt-Affected Agricultural Lands: Feasibility and Influence of Soil Properties

Abstract

Mapping Soil Organic Carbon (SOC) at a regional scale is essential for assessing soil health and supporting sustainable land management. This study evaluates the potential of using Sentinel-2 imagery and regional calibration to predict SOC in salt-affected agricultural lands in Portugal while also assessing the influence of soil properties, such as texture and salinity, on SOC prediction. A per-pixel mosaicking approach was set to analyze the relationship of spectral reflectance indices linked to bare soil conditions with SOC. SOC prediction models were developed using linear regression (LR) and Partial Least Squares Regression (PLSR). Among the tested approaches, the combination of the maximum Bare Soil Index (maxBSI) with LR produced the most accurate SOC predictions, achieving moderate prediction performance (R² = 0.52; RMSE = 0.16%; LCCC = 70%). This approach slightly outperformed the application of the 90th percentile of bare soil pixels (R90 reflectance) and the median approaches with PLSR. Notably, our findings indicate that soil salinity did not significantly affect SOC predictions within the observed salinity range of EC_e between 1.2 and 10.4 dS m⁻¹ in topsoil. However, further case studies are needed to validate this observation across diverse agricultural conditions. In contrast, soil texture and moisture content emerged as the dominant factors influencing soil reflectance. The combination of per-pixel mosaicking and regional calibration provides a practical, scalable, and cost-effective method for generating SOC maps using open access satellite imagery. To support wider adoption and improve model generalizability, future studies should incorporate a larger number of fields with a wider range of soil properties, crop types, and management practices.

1. Introduction

Soil Organic Carbon (SOC) is one of the most critical indicators of soil health and fertility, playing an important role in the productivity and sustainability of agricultural ecosystems. SOC contributes to soil structure, water retention, nutrient cycling, and microbial activity, all of which are essential for promoting crop yield and soil resilience against environmental stresses such as drought and erosion [1,2]. Along with its role in the adaptation of agricultural systems to climate change, SOC can play a role in climate mitigation by sequestering atmospheric carbon [3]. The assessment of and increase in or conservation of SOC are therefore key priorities for both farmers and policymakers who seek to enhance agricultural productivity while preserving ecosystem health. However, assessing SOC over large spatial scales presents significant challenges. Traditional soil sampling and laboratory analyses are resource-intensive and time-consuming, particularly when applied across large agricultural regions.

Recent advancements in remote sensing technologies, particularly the availability of high-resolution satellite data from missions like Sentinel-2, have opened new opportunities for the large-scale assessment of soil properties, including SOC. Sentinel-2, a satellite mission launched under the Copernicus program of the European Space Agency (ESA), provides high-resolution and multispectral imagery that allows for the large-scale and frequent monitoring of land surfaces, making it possible to assess soil properties like SOC at a more refined spatial and temporal resolution than earlier satellite constellations, such as Landsat. Various studies have already analyzed the utility of Sentinel-2 data for SOC assessment [4,5,6,7]. However, the need to account for other soil properties’ impact on SOC prediction from Sentinel-2 imagery was highlighted (e.g., [8]), as these parameters can significantly affect the accuracy of SOC predictions. In one such study, Castaldi et al. [9] used Sentinel-2 time series of several spectral indices and modeling approaches to predict SOC in ten agricultural fields. The authors found that the approaches with the best results were not the same for every field.

Soil texture variability poses a significant challenge in predicting SOC using multispectral satellite data [10]. The role of soil texture in estimating SOC using multispectral satellite imagery can be either beneficial or limiting, depending on how these variables are spatially related [11]. For remote sensing applications aimed at predicting SOC, it is important to consider how soil texture and organic matter content are correlated. When SOC and soil texture follow similar or opposing spatial patterns, their combined influence on surface reflectance may enhance prediction outcomes. On the other hand, if their spatial variability is independent, soil texture effects may interfere with the spectral signals related to SOC, potentially diminishing model reliability [11]. Both SOC and finer soil particles also tend to enhance the soil’s capacity to retain water, which consequently impacts reflectance due to absorption at key wavelengths [12]. However, the relationship between SOC and soil texture components is inherently complex and influenced by multiple factors, including soil type, aggregation, land use, and management practices [13].

Collaborative efforts have been made to map agricultural topsoil SOC and further explore how various soil properties influence SOC predictions across diverse agricultural fields in European countries within the STEROPES project, funded under the European Joint Program (EJP SOIL). As a part of this joint initiative, this study builds on previous findings by further examining the relationship between SOC and soil texture while also incorporating soil salinity, an overlooked factor in research but particularly relevant in coastal and salt-affected agricultural regions. Soil salinity, which alters the soil’s physical and chemical characteristics, can also influence the spectral response, adding further complexity to SOC assessment in coastal and salt-affected agricultural regions.

An increase in soil salinity has been shown to negatively affect SOC. For instance, Hassani et al. [14] showed that even modest increases in salinity can reduce SOC levels. Importantly and relevant to this study, the magnitude of this impact also varies with soil texture [15]. In a controlled laboratory study, She et al. [16] reported that coarse-textured soils (e.g., sandy loam and sandy clay loam) were more adversely affected by salinity than fine-textured soils (e.g., silty clay). This soil texture-dependent response to salinity introduces an additional layer of complexity when assessing the influence of soil texture on SOC prediction derived from multispectral satellite data, as explored in previous studies.

Another aspect of this study involves evaluating the possibility of establishing a regional calibration approach based on the samples from a limited number of strategically selected fields, with expected contrasting soil properties and crop types, representative of expected soil properties and crop type variabilities across the region. This limited-but-representative sampling approach enables the potential use of the resulting regression model to predict soil properties not only within the sampled fields but also in surrounding areas across the same region [17]. An added advantage is that the samples collected from fields with high soil variability provide a broad range of soil property data, facilitating regression between remote sensing data and soil properties from a limited field number.

To achieve this aim, we used “per-pixel” mosaicking approaches based on calculating the reflectance and indices related to bare soil and the driest soil conditions, as described in Vaudour et al. [18]. These approaches are advantageous compared to “per-date” approaches [18] or selecting a specific date based on soil conditions [9,19] in our case study for two main reasons: (i) no singular date exhibited bare soil conditions across all fields, and (ii) these approaches allow for the potential use of the resulting regression model to predict SOC and other soil properties not only within the sampled fields but also in surrounding areas across the same region.

2. Materials and Methods

2.1. Study Area

Coastal soils in Portugal support intensive agricultural production, sustain local economies, and serve as critical habitats for migratory birds. However, these regions face a high risk of soil salinization [20]. Our study area is Lezíria de Vila Franca, a peninsula of alluvial origin surrounded by the Tagus and Sorraia rivers, located 10 km northeast of Lisbon, Portugal (Figure 1). The area is of a significant agronomic, ecological, and socio-economic value, but it is increasingly vulnerable to salinity due to its proximity to the Tagus River estuary. The influence of estuarine tides on groundwater contributes to soil salinity, threatening soil health and productivity. Previous studies have identified salinity levels ranging from moderate to extreme across the study area. Soil salinity is higher in the south as a result of the largest influence of saline water of the estuary on the groundwater in the southern part of the peninsula [21,22]. The region is a hot-summer Mediterranean climate (Csa), as classified by the Köppen–Geiger system [23]. The soils are classified as Fluvisols [24]. The area consists of agricultural fields cultivated with annual crops, irrigated through a pressurized network that uses surface water from the Tagus River. After harvesting the main crop, residues are left on the ground, and a cover crop is seeded soon thereafter, making bare soil conditions rare.

2.2. Ground-Truth Data

Four agricultural fields, characterized by different cropping systems and soil salinity levels, were selected for soil sampling. Figure 2 shows the locations of these fields, the sampling points, and the dominant land uses across the study area. In 2022, the primary land uses were rice and tomato, with rice fields predominantly located in the central and southern parts of the area and tomato fields mainly in the north. Maize was the third most cultivated crop, although its distribution was more scattered, primarily in the northern and eastern parts. Several fields were also cultivated with other crops such as potato, sunflower, broccoli, and bell pepper, though these were very few fields within the study area. The remaining areas, which were not classified, were not under cultivation in 2022. This primarily includes the southern part of the study area, which is currently unsuitable for agriculture due to high salinity levels.

A total of 63 composite soil samples were collected at a depth of 0–20 cm in a grid pattern across the four fields in October 2022 and March 2023. The number of collected soil samples from each field was adjusted based on field conditions, and the sampling dates occurred after the harvesting of the main crop. Fields A and B were planted with tomatoes, with tomato residues left on the soil. Field C was planted with maize, and some maize stubbles were left in the soil. Field D was a rice field, and rice residues were incorporated into the soil.

Soil samples were analyzed for SOC, the electrical conductivity of the saturated soil paste extract (EC_e), particle size distribution, pH, and gravimetric water content (Ө_g). SOC was determined using the Walkley–Black colorimetric method [25], and EC_e was measured in the extract of the saturated soil paste [26]. The particle size distribution was determined using the pipette method for particles smaller than 20 µm (clay and silt fractions) and by sieving for particles ranging from 20 to 2000 µm (fine and coarse sand fractions), following the international classification system of the International Union of Soil Sciences [27]. pH was measured in a suspension of a 1:5 volume fraction of soil in water [28].

2.3. Satellite Spectral Data Selection

In the Google Earth Engine (GEE) platform, satellite data from the Copernicus Sentinel-2 Multi-Spectral Instrument (MSI) at Level 2A was selected. The MSI comprises 13 spectral bands that encompass the visible, near-infrared (NIR), and shortwave infrared (SWIR) regions. Of these, four bands have a 10 m spatial resolution (Blue, Green, Red, and near-infrared), while six bands offer a 20 m resolution, including red edge, SWIR, and additional NIR bands, and three bands have a 60 m spatial resolution (coastal blue, water vapor, and cirrus). The details of the bands used in this study are shown in

Table 1. (a) The features of the spectral bands used in this study. (b) The spectral bands/indices and modeling methods used in the six per-pixel mosaicking approaches.

(a)
Spectral Band	Spatial Resolution (m)	Central Wavelength (nm)	Band Width (nm)
B2 (Blue—B)	10	490	65
B3 (Green—G)	10	560	35
B4 (Red—R)	10	665	30
B5 (Red edge—RE1)	20	705	15
B6 (Red edge—RE2)	20	740	15
B7 (Red edge—RE3)	20	783	20
B8 (Near-infrared—NIR)	10	842	115
B8A (Narrow near-infrared—NIRN)	20	865	20
B11 (Shortwave infrared—SWIR1)	20	1610	90
B12 (Shortwave infrared—SWIR2)	20	2190	180
(b)
Mosaicking Approach	Spectral Bands/Index		Modeling Approach
Median	B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12		PLSR
R90	B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12		PLSR
MaxBSI	$BSI={\displaystyle \frac{\left(B12+B4\right)-(B8+B2)}{\left(B12+B4\right)+(B8+B2)}}$		LR
MinS2WI	$S2WI={\displaystyle \frac{B8A-B11-B12}{B8A+B11+B12}}$		LR
MinNDVI	$NDVI={\displaystyle \frac{B8-B4}{B8+B4}}$		LR
MinNDI	$NDI={\displaystyle \frac{B2-B12}{B2+B12}}$		LR

a. The Level 2A product, provided by ESA, offers Bottom of Atmosphere (BOA) reflectance images derived from Level 1C data, processed through the Sen2Cor atmospheric correction tool (L2A) [29].

The selected time period for this study was defined as one year, spanning from 1 June 2022 to 1 June 2023, to implement a per-pixel mosaicking approach, and images with cloud cover above 10% were previously excluded. Three spectral indices were then calculated for each pixel to exclude pixels with non-bare soil conditions: The Normalized Difference Vegetation Index (NDVI) (Equation (1)), the Normalized Burn Ratio 2 (NBR2) (Equation (2)), and the Bare Soil Index (BSI) (Equation (3)).

$$NDVI={\displaystyle \frac{B8-B4}{B8+B4}}$$

(1)

$$NBR2={\displaystyle \frac{B11-B12}{B11+B12}}$$

(2)

$$BSI={\displaystyle \frac{\left(B12+B4\right)-(B8+B2)}{\left(B12+B4\right)+(B8+B2)}}$$

(3)

The combined use of these indices has been found to be effective in distinguishing bare soil from other land covers. The pixels with pure bare soil conditions were obtained by selecting pixels with NDVI below 0.35, NBR2 below 0.125, and BSI over 0.021, as proposed by Castaldi et al. [9].

After filtering pixels on each image based on these criteria, six per-pixel mosaicking approaches were implemented to determine the optimal bare soil conditions across the study area within the one-year time series. The following six approaches, presented in Table 1b, were employed:

• Median: The median reflectance value for each band was calculated throughout the time series, aiming to represent typical bare soil conditions by minimizing the influence of extreme reflectance values. All ten bands were used in this approach. This approach ensures spectral data that is less affected by anomalies.
• R90: The reflectance values at the 90th percentile were selected for each band throughout the time series to identify drier soil conditions, assuming that higher reflectance values correspond to lower soil moisture. All ten bands were used in this approach. Using the 90th percentile instead of only the maximum reflectance helps eliminate anomalous reflections caused by imperfect cloud masking [9].
• MaxBSI: Pure bare soil conditions were identified by selecting the date with the maximum BSI for each pixel over the time series, using Equation (3).
• MinS2WI: Reflectance data corresponding to the date with the minimum Soil Water Index (S2WI) (Equation (4)) were selected to target the driest soil conditions for each pixel across the time series.

$$S2WI={\displaystyle \frac{B8A-B11-B12}{B8A+B11+B12}}$$

(4)

5.

MinNDVI: To capture conditions with minimal vegetation, the date with the lowest NDVI was identified for each pixel over the time series, using Equation (1).

6.

MinNDI: Various “normalized difference index—NDI”—values apart from scenarios 3–5 were tested, and the minimal NDI of the following index (Equation (5)) was selected as the best NDI in our case study.

$$NDI={\displaystyle \frac{B2-B12}{B2+B12}}$$

(5)

2.4. Modeling Approaches for SOC Prediction

The regional calibration approach is designed to develop predictive models for soil properties using data from a limited number of sampling sites with contrasting soil properties. This approach is advantageous over field-specific calibration as it allows for the application of the resulting regression model across a broader region, making it practical for areas with high soil variability and numerous small, privately owned farms. The methodology begins with the selection of sampling fields that represent the expected range of soil conditions and crops within the study area. Soil samples were collected from these fields and analyzed for various properties, as detailed in Section 2.2. High-resolution satellite data were then used to develop regression models that relate spectral values and indices (derived using six distinct per-pixel approaches in GEE, as detailed in Section 2.3) to the measured soil properties. Spectral values and indices were extracted at each sampling point location using the sampleRegions function in GEE. This function extracts the pixel values of each resultant image composite or derived index at the location of each point geometry.

Partial Least Squares Regression (PLSR) was employed for the first two scenarios to develop predictive models for soil properties based on bare soil sample spectra. PLSR relates the explanatory variable matrix (in this study, soil reflection indices) to the dependent variable matrix (in this study, soil properties) through a linear multivariate approach, leveraging latent variables to address noise and collinearity in the data [30,31]. PLSR is a widely utilized method in remote sensing applications due to its ability to reveal the impact of spectral bands on various soil properties prediction [8]. Additionally, compared to machine learning methods (e.g., Random Forest, Support Vector Machine), PLSR is usually more robust and less susceptible to overfitting, particularly when working with small datasets, which is the case in our study with a total number of 63 samples. The spectral data and outliers were identified using the Mahalanobis distance and multivariate outlier detection techniques [32]. Linear regression (LR) was applied for the remaining four scenarios to establish a regression between each index and soil properties, as a linear relationship between the indices and soil properties was predominant. Outlier removal before LR modeling followed the Z-score-based filtering approach, removing extreme values that could disproportionately influence the model’s parameters.

The validation of the models was performed using Leave-One-Out Cross-Validation (LOOCV). This method systematically uses each sample for validation, while the remaining samples serve as the training set. This procedure is iteratively repeated for each sample until each of the samples is removed once. The predictive performance of the calibrations was assessed using the root mean square error (RMSE), the residual prediction deviation (RPD), the ratio of performance to interquartile distance (RPIQ), Lin’s concordance correlation coefficient (LCCC), and the coefficient of determination (R²) between the observed and predicted values. The RMSE quantifies prediction errors by taking the square root of the mean squared differences between observed and predicted values. Lower RMSE values signify higher accuracy in the predictions. RPD reflects the ratio of the standard deviation of observed values to prediction error. Values of 1.4 to 1.8 denote models with moderate levels of predictive capacity; values of between 1 and 1.4 indicate models exhibiting poor levels of predictive capacity; and values of less than 1 denote very poor models that should not be used [19]. The RPIQ is the ratio of the interquartile range of the observed values to the RMSE of prediction [9]. LCCC evaluates the degree of agreement between observed and predicted values, measuring how well predictions align with a 1:1 relationship. Its value ranges from −1 to 1, where 1 represents perfect agreement [33]. The prediction agreement can be categorized according to LCCC as excellent (>0.9), good (0.8–0.9), moderate (0.65–0.8), or poor (<0.65) [34]. The PLSR and LR models were implemented using Python’s Scikit-learn library (version 1.0.2), with the number of latent variables (NL_v) for the PLSR approach optimized during the cross-validation process based on the RMSE. The final number of samples included in the PLSR and LR analysis was determined after removing outliers.

Furthermore, to evaluate the model’s uncertainty estimates, the proportion of cross-validation observations falling within the 90% prediction interval, referred to as the prediction interval coverage probability (PICP), was computed. An ideal PICP value is close to 0.9 (i.e., 0.88–0.92), indicating that the model’s uncertainty estimates are well-calibrated. Values significantly above 0.9 suggest that the prediction intervals are too wide, implying an overestimation of uncertainty, whereas values significantly below 0.9 indicate overly narrow intervals and thus an underestimation of uncertainty [35]. To visualize the uncertainty of SOC prediction as a map, we calculated the prediction interval ratio, PIR [36]:

PIR = (P95 − P05)/P50

(6)

where P05, P50 (median), and P95 are the 5th, 50th, and 95th percentiles of prediction.

3. Results

3.1. The Bare Area Mapped Across the Time Series

The total number of L2A images, with a cloud cover of less than 10%, acquired for the study area over the selected one-year period is shown in Figure 3a. The map reveals significant spatial variability in the number of images. The northeastern part of the study area recorded a minimum of 34 images, while the number increased sharply toward the west, reaching over 80 images in the western zone. Notably, two small areas in the northwestern and southeastern parts of the study area exhibit an exceptionally high number of images, exceeding 100. Fields A and B are located in a zone with 34 total images, while fields C and D have a higher total of 55 images. The unexpectedly high number of images and large spatial variability across such a small study area are primarily due to the study area being located at the overlap of multiple Sentinel-2 tiles, resulting in some areas being covered by a greater number of images.

Figure 3b presents the total number of bare soil conditions for each pixel, obtained after applying the filtering criteria for the NDVI, NBR2, and BSI, as described in Section 2.3. The map indicates that only a very small part of the study area (marked in white) was never detected with bare soil throughout the time series. These zones mostly correspond to local access roads within the agricultural land. The spatial variability in the number of bare soil conditions per pixel largely follows the same trend observed for the L2A images, with a higher number of bare soil conditions in the western part of the study area. Regarding the four fields, all have at least 10 bare soil dates per pixel. Field A has the highest number of bare soil conditions (18), followed by fields B and D with 12 images each, while field C has 10 bare soil conditions per pixel. The smallest numbers of bare soil conditions per pixel were found in the east-central part of the study area, likely due to the lower overall availability of images in the eastern section and the predominant land use for rice cultivation.

3.2. Analysis of Soil Properties

Table 2. Characteristics and basic statistics for soil properties for the four sampled fields.

Field	Date of Sampling	Soil Cover	Area (km2)	Nº of Samples	SOC [%]				Clay [%]				Sand [%]
					Min	Max	Mean	Std	Min	Max	Mean	Std	Min	Max	Mean	Std
A	Oct 2022	Tomato residues	23	20	0.9	1.4	1.2	0.1	30.1	44.9	35.5	4.9	13.6	36.2	24.6	6.3
B	Oct 2022	Tomato residues	33	14	0.9	1.5	1.2	0.2	37.0	46.3	41.9	2.6	12.5	22.6	16.0	2.9
C	Oct 2022	Maize stubble residues	34	13	1.2	1.8	1.4	0.2	35.6	52.4	49.4	4.6	5.4	9.9	7.0	1.27
D	Mar 2023	Rice residues	3	16	1.3	1.8	1.5	0.2	47.1	50.8	49.1	1.1	5.7	7.2	6.5	0.5
All				63	0.9	1.8	1.3	0.2	30.1	52.4	43.2	7.0	5.4	36.2	14.5	8.7
Field	Date of Sampling	Soil Cover	Area (km2)	Nº of Samples	θg [%]				Ece [dSm−1]				pH 1:5
					Min	Max	Mean	Std	Min	Max	Mean	Std	Min	Max	Mean	Std
A	Oct 2022	Tomato residues	23	20	10.3	19.5	14.8	2.5	1.6	6.2	3.1	1.3	5.8	7.6	6.8	0.4
B	Oct 2022	Tomato residues	33	14	11.1	26.5	19.2	4.7	2.2	10.4	5.1	2.4	6.8	8.5	7.6	0.5
C	Oct 2022	Maize stubble residues	34	13	19.0	27.3	24.4	2.3	1.8	4.4	3.0	0.7	7.6	8.5	8.2	0.3
D	Mar 2023	Rice residues	3	16	38.0	55.1	45.4	5.7	1.2	1.9	1.6	0.2	7.7	8.2	8.1	0.1
All				63	10.3	55.1	25.5	12.8	1.2	10.4	3.1	1.8	5.8	8.5	7.7	0.7

presents the statistics for SOC, clay and sand content, EC_e, and pH for the four fields. The SOC content varies within a narrow range (0.9–1.8%), indicating relatively low organic carbon levels across the study area. The highest mean SOC value (1.5%) is observed in field D (rice), while the lowest (1.2%) is found in fields A and B (tomato). The small variation in SOC can be attributed to the soil type, as Fluvisols in Mediterranean climates tend to have lower SOC due to higher mineralization rates. It may also be influenced by similar cropping systems, residue management practices, and soil moisture conditions, as all fields are irrigated. The clay content ranges from 30.1% to 52.4%, with field A having the lowest clay content and fields C and D the highest. In contrast, field A shows the highest sand content with an average of 24.6%, followed by relatively lower average value of 16% in field B and the lowest average sand contents of 7% and 6.5%, respectively, in fields C and D. Field A shows the largest within-field variability in soil texture with sand and clay content in the ranges of 13.6–36.2% and 30.1–44.9%, respectively.

The gravimetric water content (θ_g) varies significantly across fields, ranging from 10.3% to 55.1%. Field D exhibits the highest water contents (mean 45.4%), which is expected due to likely prolonged soil saturation. In contrast, fields A and B (tomato residues) show lower moisture levels, with means of 14.8% and 19.2%, respectively. Field C (maize stubble residues) presents intermediate moisture values (mean 24.4%). EC_e, a key indicator of soil salinity, varies significantly, from 1.2 to 10.4 dS m⁻¹, across the study area. Field B exhibits the highest EC_e values (mean 5.1 dS m⁻¹, max 10.4 dS m⁻¹), revealing moderate to high soil salinity. In contrast, field D has the lowest EC_e values (mean 1.6 dS m⁻¹), indicating non-saline soil conditions at the time of sampling. Fields A and C present low- to moderate-salinity conditions with mean values of 3 and 3.1 dS m⁻¹ and max values of 6.2 and 4.4 dS m⁻¹, respectively. Soil pH varies between 5.8 and 8.5, with field A showing the lowest pH (mean 6.8), while fields C and D have the highest (mean 8.1–8.2), indicating slightly alkaline soil conditions.

On a regional scale, considering the total variability in soil properties across the four fields, SOC exhibits relatively low variability, ranging from 0.9% to 1.8%, with an average of 1.3% and a small standard deviation, similar to the values observed within each field. In contrast, clay and sand content display significantly greater variability at the regional level, with larger standard deviations, suggesting that soil texture may have a stronger influence on bare soil reflectance. θ_g also shows considerable spatial differences, with high variability across fields, reinforcing the strong relationship between soil texture, organic matter, and moisture retention capacity. Regarding EC_e, field B shows the highest salinity variability, with a relatively larger standard deviation compared to regional variability. On the other hand, pH exhibits greater variability at the regional scale, following the trend observed for clay and sand content.

Figure 4 presents the Spearman correlation coefficients between measured soil properties, with significant relationships at p < 0.01. The results indicate a good correlation between SOC and clay content and θ_g, indicating the well-established link between fine-textured soils and higher organic carbon retention. In addition, a moderate correlation exists between SOC and pH, suggesting that, overall, the higher SOC content was found in more alkaline soil. On the other hand, the negative and weak correlation between SOC and EC_e indicates overall lower SOC in more saline topsoil.

This negative soil salinity–SOC relationship is further supported by a recent large-scale study by Hassani et al. [14], who analyzed 43,459 topsoil samples across different land covers since 1992. They observed that increasing salinity from 1 to 5 dS m⁻¹ corresponded to a decline in SOC by approximately 4.4% in croplands and 9.3% in non-croplands. Although salinity explained a modest portion of SOC variance (<6%) in their study, these results empirically confirm that even modest soil salinity increases can impact SOC.

3.3. Performance of Prediction Models for Soil Properties

Table 3. Cross-validation performances (a) for median and 90th percentile reflectance values using PLSR method (b) for various indices using LR approach. For each property, the model with the best prediction performance is highlighted.

(a)
Soil Property					Median														R90
	RMSE			R2	LCCC		RPD			RPIQ	NLV				RMSE		R2		LCCC		RPD			RPIQ		NLV
SOC	0.17			0.44	0.63		1.35			1.61	3				0.16		0.50		0.68		1.42			1.74		2
Clay	3.13			0.80	0.89		2.24			3.85	3				3.43		0.76		0.87		2.04			3.50		4
Sand	3.40			0.84	0.92		2.55			4.18	3				3.59		0.83		0.91		2.43			3.96		5
θg	5.03			0.84	0.91		2.53			2.88	4				5.99		0.78		0.88		2.14			2.42		6
ECe	1.51			0.31	0.51		1.21			1.33	3				1.55		0.29		0.49		1.19			1.31		3
pH	0.41			0.66	0.80		1.74			3.01	3				0.35		0.75		0.86		2.01			3.27		2
(b)
Soil Property	maxBSI						minS2WI							minNDVI								minNDI
	RMSE	R2	LCCC	RPD		RPIQ	RMSE	R2	LCCC			RPD	RPIQ	RMSE		R2		LCCC	RPD	RPIQ		RMSE	R2		LCCC	RPD	RPIQ
SOC	0.16	0.52	0.70	1.45		1.79	0.17	0.41	0.60			1.31	1.62	0.19		0.28		0.46	1.18	1.46		0.16	0.50		0.69	1.43	1.77
Clay	3.94	0.68	0.83	1.80		3.10	5.04	0.48	0.67			1.40	2.42	5.35		0.41		0.61	1.32	2.28		4.09	0.66		0.81	1.72	2.99
Sand	4.06	0.76	0.88	2.05		3.30	5.58	0.55	0.72			1.50	2.40	5.81		0.51		0.69	1.43	2.31		4.09	0.76		0.88	2.04	3.27
θg	5.02	0.84	0.93	2.55		3.43	5.05	0.84	0.93			2.53	3.41	9.99		0.38		0.57	1.28	1.72		5.31	0.83		0.92	2.41	3.25
ECe	1.45	0.15	0.29	1.12		1.32	1.20	0.29	0.47			1.19	1.41	1.34		0.11		0.24	1.07	1.26		1.29	0.18		0.34	1.11	1.31
pH	0.45	0.53	0.71	1.46		2.43	0.52	0.37	0.57			1.27	2.12	0.45		0.54		0.72	1.48	2.47		0.44	0.56		0.73	1.51	2.52

shows the cross-validation results of the models developed for the prediction of the six soil properties, using the six considered approaches. PLSR was used with the median and R90 approaches, and LR was used with maxBSI, minS2WI, minNDVI, and minNDI. For each property, the model with the best prediction performance is highlighted.

In terms of SOC, both the median and R90 approaches show moderate prediction ability, with the R90 approach providing slightly better overall performance (RMSE = 0.16%; RPD = 1.42; RPIQ = 1.74; R² = 0.50) compared to the median approach (RMSE = 0.17%; RPD = 1.35; RPIQ = 1.61; R² = 0.44). Additionally, LCCC indicates improved model agreement for the R90 approach. For clay content, both methods show very good prediction ability; however, the median model performs slightly better compared to R90. Similarly, sand prediction is marginally better with the median approach than R90, although both approaches present very good prediction ability. For θ_g, both approaches demonstrate strong prediction ability, with the median approach achieving slightly better prediction ability. Both models maintain high agreement for clay, sand, and θ_g predictions, as indicated by LCCC values, all above 0.85. In contrast, both models show weak predictive capabilities for EC_e, with RMSE values around 1.5 dS.m⁻¹ and R² below 0.35. However, pH predictions show stronger performance, particularly using the R90 approach.

The cross-validation results for the LR models using different spectral indices (Table 3b) show that SOC prediction performs the best using maxBSI (RMSE = 0.16%; RPD = 1.45; RPIQ = 1.79; R² = 0.52), closely followed by minNDI (RMSE = 0.16%; RPD = 1.43; RPIQ = 1.77; R² = 0.50). Clay and sand predictions also show better results with maxBSI and minNDI compared to minS2WI and minNDVI. For θ_g, the best performance is observed using maxBSI and minS2WI, achieving high predictive accuracy. In contrast, minNDVI yields significantly weaker results, while minNDI provides slightly poorer but still strong performance. Notably, EC_e prediction remains weak across all indices, with R² and RPD values below 0.30 and 1.40, respectively. pH predictions perform relatively well across all indices, except for minS2WI, with an R² below 0.40 and RPD below 1.40.

A comparison of the cross-validation results highlights differences in predictive performance across soil properties. PLSR models generally outperform LR models for most soil properties (i.e., clay, sand, and pH), while SOC and θ_g show better performance in the LR model when using maxBSI. This suggests that no single model is optimal for predicting all selected soil properties. Nevertheless, none of the presented models shows a good prediction ability for EC_e assessment.

3.4. SOC Prediction

The results from Table 3 indicate that the maxBSI approach provides the best prediction performance (R² = 0.52; RMSE = 0.16%; LCCC = 0.70) among all the tested approaches. Given that LR is computationally less demanding than PLSR while also showing slightly better performance for SOC prediction, this model was selected for SOC prediction.

Figure 5 presents the maxBSI map of the study area, where maxBSI values range from 0.021 to 0.2. As described in Section 2.3, pixels with BSI values lower than 0.021 were excluded, ensuring that only dates with pixels with bare soil conditions were considered. Analyzing this map reveals that maxBSI is generally higher in the northern part of the study area, which aligns with the known north-to-south soil texture variability. The northern region is characterized by sandier soils, as reflected in our soil texture analysis from the four sampled fields (Table 2), leading to brighter soil reflectance. In contrast, lower maxBSI values are concentrated in the central part of the study area, where rice cultivation dominates. The lower BSI in this region could be attributed to higher clay content, wetter soil conditions, and higher organic carbon levels, all of which contribute to darker soil reflectance. Notably, the southern part of the study area also exhibits slightly higher maxBSI values compared to the center. While high clay content is expected in this region, most of this region is not used for agriculture (see Figure 2), due to high salinity levels. The slightly higher maxBSI values in these southern non-agricultural areas may be attributed to drier soil conditions, lower organic matter content, more exposed mineral soil, and potentially salt-affected topsoil that appears brighter than clayey agricultural fields.

In terms of maxBSI variability with the sampling fields, field A presents the highest maxBSI values with the largest interfield variability, ranging from 0.12 to 0.17, followed by field B with the second largest maxBSI and relatively lower variability in the 0.11–0.13 range. Field C shows a lower maxBSI in the range of 0.09–0.11, while field D presents the field with the lowest maxBSI, with values in the 0.05–0.07 range. The spatial variability in maxBSI within the sampling locations agrees with the already known overall variability in soil texture, with sandier soil in the north to more clay-rich soils in the center and south. The lowest maxBSI in field D is also not unexpected due to the significantly larger moisture content in rice fields. The total variability in maxBSI within these four sampled locations (0.05–0.17) covers the observed total variability in maxBSI across the study area relatively well, indicating that the selected four fields represent the variability in maxBSI across the study area well.

Figure 6a shows a scatter plot between maxBSI and measured SOC in the four fields, along with the LR model used for the prediction of SOC:

SOC = 1.84 − 4.09526 maxBSI

(7)

Figure 6b shows the relationship between predicted and measured SOC across the four fields using the LR model. The color-coded labels help visualize the contribution of each field and evaluate the potential for field-specific calibration within the study area. Upon closer inspection, the relatively low within-field variability in both maxBSI and measured SOC appears to limit the feasibility of field-specific calibration, as no significant correlation is observed between these parameters at the field scale. This indicates that while the model may not provide sufficient precision for capturing within-field SOC variability, it offers acceptable predictive accuracy at the regional scale for between-field SOC assessment, based on the cross-validation performance detailed in Table 3b.

3.5. Mapping SOC in the Study Area

A map with the predicted SOC of the study area is presented in Figure 7. The map shows variations in SOC in the 0.75–1.75% range, representing relatively small variability. Areas with low BSI values (typically associated with high soil moisture and clay content) likely correspond to higher-SOC regions, while high BSI values (indicative of dry and sandy soils) correspond to lower-SOC-content zones. The general SOC content variability gradient from north to south aligns with the expected soil texture gradient, where the northern region has higher sand and lower clay content, transitioning to the central and southern region, which has lower sand and higher clay content.

The highest SOC values were primarily found in rice-growing areas, where higher SOC content was anticipated. This pattern was also observed in our sampling location within the rice field (field D), which exhibited slightly higher SOC content. This difference can be partially attributed to the flooded conditions in rice fields, which reduce oxygen availability, slow down organic matter decomposition, and promote SOC accumulation. In contrast, tomato and maize fields are generally located in areas with lower clay content, which may also contribute to lower SOC levels due to the reduced stabilization of organic matter.

It is worth noting that the regional calibration approach may have limited ability to predict within-field SOC variability, due to the relatively low variability observed within fields for both maxBSI and measured SOC, as discussed in Section 3.4 (interpretation of Figure 6). This low variability, combined with the lack of a significant correlation between these parameters at the field scale, limits the model’s ability to accurately capture within-field SOC variations.

3.6. Uncertainty Map

To assess the spatial distribution of model uncertainty, a PIR map was generated using Equation (5) and presented in Figure 8. The resulting PIR values ranged from 0.29 to 0.68 across the study area, highlighting areas of varying levels of relative uncertainty in the SOC predictions. A clear spatial pattern emerged, with higher PIR values predominantly located in regions where predicted SOC was lower, particularly in areas with SOC lower than 1.0%. This observation aligns with the theoretical expectation that when the predicted value is smaller, even small absolute prediction intervals result in a larger relative uncertainty, as reflected in the PIR calculation. In contrast, areas with higher SOC values (above 1.35%) exhibited lower PIR values, suggesting greater model confidence in predictions for these regions. Higher PIR values in areas of lower SOC indicate that these regions are more challenging for the model to predict with confidence. This may be due to inherent model limitations in regions with lower organic carbon content and possibly worse-calibrated data in this area due to a smaller number of samples in the SOC range of lower than 1%. Conversely, lower PIR values in areas of higher SOC suggest that the model performs more reliably in regions with higher levels of SOC.

For a 90% prediction interval, we observed a PICP of 0.92, indicating that the obtained prediction intervals are a realistic representation of the prediction uncertainty with a slight overestimation of uncertainty, as the expected value for a 90 % prediction interval is 0.90.

4. Discussion

The following discussion focuses on three key aspects of this research: the influence of soil properties on SOC prediction and the strengths and limitations of the proposed approach.

4.1. Influence of Soil Properties on SOC Prediction from maxBSI

To investigate the influence of other soil properties on maxBSI, examine how the correlation between SOC and other soil properties affects SOC prediction from maxBSI, and provide better insight into the regional vs. field-specific calibration approach, we plotted maxBSI against the soil parameters described in Table 2 in Figure 9. A detailed examination of Figure 9 suggests the following:

• Soil texture (clay and sand content) exhibits a stronger correlation with maxBSI than with SOC, consistent with the results in Table 3. We attribute this to the greater spatial variability in soil texture across the study area, making it the dominant factor influencing soil reflectance. However, since SOC is correlated with clay content in this region, this relationship likely contributes to the ability to predict SOC from bare soil images with moderate accuracy. These findings align with the recent study of Wetterlind et al. [11], across 34 fields in 10 European countries, which found improved SOC prediction performance with increasing SOC-to-clay correlation. The observed significant correlation between clay and pH also suggests that pH can be predicted from bare soil images, as expected.
• θ_g shows the strongest correlation with maxBSI among all soil properties. Although soil moisture is highly dynamic and cannot be directly compared to the maxBSI derived from a per-pixel mosaicking approach, the strong observed correlation is likely due to the systematic spatial variability in soil moisture across the study area. In particular, the northern fields, which have sandier soils, tend to retain less moisture, a pattern that remains consistent during bare soil conditions. Additionally, while a one-year time series was used in this approach, images with bare soil conditions are mostly concentrated within a shorter period of the year. The strong correlation between θ_g and maxBSI can also be explained by the substantially higher moisture content in the rice field, where prolonged soil saturation results in darker soil, further affecting maxBSI. This effect appears to have a greater impact than clay content, as fields C and D have comparable clay levels (see Table 2).
• EC_e showed no significant correlation with satellite data, considering the six tested approaches. Although previous studies have reported much higher soil salinity levels in the area, with EC_e exceeding 32 dS/m at deeper layers due to the influence of saline groundwater (Farzamian et al. [17]), the salinity in the topsoil, from which samples were taken in this study, was comparatively lower, ranging from 1.2 to 10.4 dS/m. A key hypothesis is that salinity levels occurring at the soil surface are not capable of significantly affecting spectral reflectance. This suggests that soil salinity in the topsoil may not significantly impact SOC prediction from bare soil imagery within the salinity range where most crops remain viable, as surface salt accumulation is likely minimal.
• Field-specific calibration is not feasible for most soil properties, consistent with the findings for SOC. This is largely due to the low within-field variability in soil properties in individual fields, which limits the statistical detectability of meaningful relationships. The only exceptions are clay and sand content in field A, which still show a strong correlation with maxBSI (Figure 9a,c). This is due to the greater variability in soil texture in field A, as previously discussed in Section 3.2 (see Table 2), whereas the other fields exhibit lower within-field variability.

Figure 9. Plots of maxBSI and measured (a) clay [%], (b) θ_g [%], (c) sand [%], (d) EC_e (dS m⁻¹), and (e) pH at all four locations.

Figure 9. Plots of maxBSI and measured (a) clay [%], (b) θ_g [%], (c) sand [%], (d) EC_e (dS m⁻¹), and (e) pH at all four locations.

4.2. The Strengths of the Proposed Approach

The “per-pixel” mosaicking approach proved particularly effective in this study. By capturing, for each pixel, the bare soil reflectance across Sentinel-2 time series, it allowed us to build a regional calibration based on samples from two sampling seasons, an especially important advantage, as rice fields were not bare during the first campaign. This method enables the inclusion of soil samples collected at different times within the selected window for the temporal mosaicking approach. In contrast, “per-date” approaches face a major limitation in heterogeneous agricultural landscapes: no single date guarantees bare soil across all farms. Consequently, the per-date approach would require field-specific calibrations and synchronized soil sampling during bare soil periods for each farm—an impractical and resource-intensive requirement.

Our findings also indicate that a regional calibration approach is a more practical solution for assessing soil properties across the study area. While field-specific calibration may offer higher precision at individual field levels, the broader applicability and efficiency of regional calibration outweigh this limitation, even in cases where field-specific calibration might be feasible. From an agronomic and soil management perspective, regional calibration is more suitable for advising farmers and irrigation managers, as it allows predictions to be applied across the entire region within the range of measured soil properties, particularly SOC, as the main focus of this study. The regional approach also aligns more effectively with the decision-making frameworks needed for the implementation of agricultural soil management practices. In contrast, field-specific calibration would require the development of new calibration equations for every new field, making it resource-intensive and less practical for addressing large-scale applications.

Although predicting soil texture was not the primary objective of this study, the presented approach demonstrated strong predictive ability for assessing soil texture across the study area. While maxBSI did not yield the best model performance compared to other tested approaches, it still demonstrated a strong predictive ability using a simple linear regression approach. This suggests that maxBSI can serve as a simple LR model yet a robust predictor of soil texture in this region.

4.3. The Limitations of the Proposed Approach

The strengths of this approach are associated with several limitations that should be acknowledged and addressed in future research:

• A key limitation of this study is the limited ability of the regional model to capture within-field variability in SOC, as evidenced by the weak correlation between maxBSI and SOC at the individual field scale (see Figure 6). This result aligns with the relatively narrow range of SOC observed within individual fields, which inherently constrains the statistical detectability of relationships between SOC and maxBSI at the field scale. Thus, our model’s predictive performance stems largely from broader between-field differences, such as variations in soil texture, moisture, and management practices, rather than spectral sensitivity to SOC at finer scales. This observation is particularly important and even limiting in contexts where detecting small differences across a single field is required for farm-level management.
• The scalability and robustness of the regional calibration approach are further constrained by the limited sample size (63 soil samples), the use of only one year of satellite imagery, and the inclusion of only four fields (two tomato, one maize, one rice), each with different crop types and management practices. This limited sample size and fields restrict the ability to evaluate how generalizable the model is and whether the approach is robust across other farms. This small sample size is also insufficient to draw firm conclusions regarding the influence of salinity on SOC prediction using maxBSI.

To address these limitations, future research should aim to incorporate multi-season satellite imagery and more extensive spatial sampling in this region, especially by expanding the number of sampled farms with a wider range of soil properties, crop types, and residue management practices. This would enable a more rigorous test of the model’s applicability to new, unseen fields and answer key questions:

(i)

Can this approach reliably assess SOC in new fields with different characteristics across the study area?

(ii)

Under what conditions, if any, does field-specific calibration become feasible?

The inclusion of more fields from the southern part of the study area, where higher topsoil salinity levels and stronger gradients are expected, would also help in better investigating the role of salinity in SOC prediction using spectral indices.

Given the small sample size in the current study, the use of PLSR was appropriate in this study, as it offers greater stability and interpretability with limited data, effectively handles multicollinearity, and highlights key spectral traits. It is a well-established method in SOC prediction and has often performed comparably to more complex models [8]. However, expanding the dataset would allow future studies to explore more advanced machine learning algorithms that generally require larger, more diverse input data. These models could improve predictive performance and provide deeper assessments of model sensitivity and variable influence.

5. Conclusions

This study highlights the potential of Sentinel-2 imagery, a per-pixel mosaicking approach, and regional calibration for predicting Soil Organic Carbon (SOC) in salt-affected agricultural lands. The key insights from this study are as follows:

• MaxBSI combined with linear regression (LR) provided the most reliable SOC predictions. The moderate prediction performance (R² = 0.52; RMSE = 0.16%; LCCC = 70%) demonstrates that while SOC can be estimated from Sentinel-2 data, its prediction accuracy is inherently constrained by the complex interactions between soil properties and spectral reflectance as well as the small range of SOC variability.
• Soil texture plays a dominant role in influencing soil reflectance, with clay and sand contents showing strong correlations with spectral indices. This explains why maxBSI, which captures bare soil reflectance variations, emerged as the best-performing index for SOC prediction. The spatial distribution of SOC across the study area followed the known north-to-south soil texture gradient, where sandier soils in the north exhibited lower SOC values.
• No strong statistical impact of soil salinity on SOC predictions was found. However, more samples are required to validate this finding across different agricultural land uses with contrasting soil properties, crop types with varying salinity tolerance, and greater variability in soil salinity to assess the method’s applicability beyond the current study area.
• The observed SOC prediction performance is driven largely by between-field differences, such as variations in soil texture and management practices, rather than within-field variability. The relatively low range of SOC within individual fields likely limits the feasibility of detecting fine-scale SOC variability at the field level in this region. As a result, field-specific calibration models show limited predictive strength and practical utility in this region.
• The regional calibration approach proved effective for SOC assessment, as the limited-but-representative sampling locations successfully represented the overall soil variability in the study area. From an applied perspective, our results also indicate that regional calibration is a more feasible and scalable approach than field-specific calibration in this study area. The relatively low within-field variability in SOC, combined with logistical challenges in accessing the large number of small private farms, limits the practicality of field-specific calibration models. Regional calibration, in contrast, allows for a broader application of the predictive model across the study area, which is particularly useful for regional-level soil monitoring, policy support, and landscape-scale SOC assessments. Nonetheless, these findings should be interpreted cautiously, as the limited number of samples and fields and their distinct characteristics may constrain the generalizability of the model. Future work should include more fields with diverse cropping systems, soil properties, and management practices to evaluate the robustness and transferability of the approach in this region.