Negative correlation between soil salinity and soil organic carbon variability

The net effect of soil salinity on the SOC content is thus a balance between the positive and negative influences. It depends on the specific interacting factors, including climate, soil type, vegetation, land use practices, and other soil physicochemical properties.

We analyzed a vast dataset comprising 43,459 soil samples from diverse biomes and land covers, particularly Europe, collected since 1992. The soil samples were collected from various depths and were obtained from well-established soil profile databases such as LUCAS (28) (n = 31,168) and WoSIS (29) (n = 12,291). These samples included soil salinity measurements, represented by soil–water extract electrical conductivity (EC) (dS m⁻¹) and SOC content (g kg⁻¹) at each sampling location. The primary objective of our research was to examine the correlation between soil salinity and SOC in mineral soils (SOC < 150 g kg⁻¹) by controlling other environmental parameters that could explain the variability in SOC. Our assumption was that a steady-state balance exists between soil, land, and climate at the sampling locations over extended periods. However, this assumption may not necessarily be valid, especially in croplands, where over short time scales land and crop management practices may play a more significant role in explaining the variability in SOC. Accordingly, we separated the analysis into croplands (n = 25,634) and noncroplands (n = 17,825). The interpretation of the developed statistical analysis enabled us to identify additional significant environmental factors and their relationship with SOC content.

To account for the complex nature of soil systems, we considered several critical environmental factors that may interact with and are correlated with the behavior of OC in the soil including climate, vegetation cover, land use practices, soil texture, soil physio-chemical properties, drainage patterns, soil moisture levels, and land management strategies. We employed general additive models (GAMs) to investigate the relation between soil salinity and SOC content while considering potential nonlinear patterns and interactions with other parameters. Given the innate correlation between these variables at large scale and the possibility of concurvity, or fundamental associations between these variables, we first ranked covariates based on their minimum redundancy and maximum relevance to SOC using MRMR (Minimum Redundancy Maximum Relevance) algorithm (30). For two major land covers, croplands, and noncroplands, we fitted separate GAMs to the SOC of soil samples with SOC < 150 g kg⁻¹, as the target variable.

In both cases, certain environmental variables were identified as the most influential in explaining SOC variability (SI Appendix, Fig. S1). Specifically, the covariates identified as significant in both croplands and noncroplands were soil total soil nitrogen (N) content (g kg⁻¹), land cover, sample depth (cm), sand content (%) in the fine earth fraction (particles < 2 mm), soil pH (measured in aqueous solution), and precipitation Seasonality Index (SI)— representing the variation in precipitation throughout the year within an area. Soil salinity emerged as one of the primary covariates in both land types. Moreover, leaf area index (LAI) was found to be significant for croplands, while topographic hill slope emerged as influential for noncroplands. One result here is that the precipitation SI emerges as a more significant covariate compared to annual precipitation and long-term air temperature.

Using GAMs allowed us to discern the specific local significance of different variables, including soil salinity, in prediction of SOC contents at varying levels of soil–water extracts (Fig. 1). For every observation, the percentage of local covariate significance for predicting SOC was calculated by dividing the absolute effect of that specific covariate by the sum of the absolute effects of all other covariates, excluding the intercept. The results indicated that the local significance of soil salinity in prediction of SOC contents ranged from 0% to ~6% across different soil-water extracts. On average, the salinity’s mean significance in prediction of SOC content was ~1.13% (SD = ~0.94%). The low mean significance of salinity suggests that other factors, such as climate, land cover, or soil texture, may have more substantial effects in prediction of SOC contents.

The analysis of the most significant local covariates in prediction of SOC values revealed soil total N, land cover, and precipitation SI, and pH among main factors highly correlated with SOC content (Fig. 2). For a total of 30,216 observations, soil N emerged as the most significant covariate, with a mean importance of ~15.35% (SD = ~8.51%). Similarly to Evans, Burke, and Lauenroth (31), our study suggests that there is a strong correlation between SOC and N levels, given that N is a main component of soil OM (32). The results align with the findings of Xu et al. (33) or Doetterl et al. (34), who highlighted the significance of geochemistry as a key factor for soil C storage, in addition to climate. Similarly, for 1,660 soil samples, land cover exhibited a notable significance (mean importance of ~5.55%, SD = ~1.38%) followed by precipitation SI, for 645 soil samples (mean importance of ~3.82%, SD = ~2.52%), suggesting the primary factors explaining SOC variability are climate and land cover type. Previous studies have also identified soil moisture, clay content, or land cover as the primary covariates (32, 35–37).

We used the developed GAMs to estimate the Accumulated Local Effects (ALEs) (38) of each covariate on the predicted SOC contents (Fig. 3). ALE plots allow understanding of how variation in each covariate influences the predicted outcome (SOC) while accounting for the interactions and nonlinearities with other covariates. We generated ALE plots for each covariate of interest. These plots display how the SOC deviates from the mean predicted SOC content as the covariate’s value ranges from its minimum to its maximum across the dataset.

Using ALEs, we found a general nonlinear decreasing trend in predicted SOC with an increase in soil salinity (Fig. 3A), for both crop and noncroplands. For croplands, ALEs are quantified by the equation ALE = 0.34 + −0.21EC + 0.01EC² while for noncroplands, ALE = −7.42 + 11.43exp(−0.31EC). In those equations, ALE represents the deviation of SOC from the mean predicted SOC (18.47 and 35.96 g kg⁻¹ for croplands and noncroplands, respectively), and EC is the soil salinity measured in dS m⁻¹. This implies that higher soil salinity is associated with reduced SOC content, consistent with geographically specific investigations of Zhang et al. (39), Qu et al. (40), Setia et al. (41), and Zhao et al. (42). The equation also provides insight into the magnitude of the correlation between soil salinity and SOC content. For example, the equation for croplands suggests that a soil salinity increase from 1 to 5 dS m⁻¹ is associated with a decrease in SOC from 0.14 g kg⁻¹ above the mean predicted SOC (18.47 g kg⁻¹) to 0.46 g kg⁻¹ below the mean predicted SOC. Such increase in soil salinity is associated with a stronger decrease of 5.95 g kg⁻¹ (ALE = 0.96 to ALE = −4.99) for noncroplands with mean SOC = 35.96 g kg⁻¹. This finding confirms that soil salinity has dominantly a negative correlation with SOC content when controlling for the role of other covariates explaining the variability of SOC. It is important, however, to note that this is a correlative analysis. We use the spatial distribution/variability of both variables to see how they are correlated. Other factors, not included in their multivariate analysis, might affect the distributions of both variables (i.e., SOC and soil salinity).

The ALEs also reveal the relations between SOC and other relevant soil and environmental parameters. The source soil profile databases also have documented a nearly linear relationship between SOC and N (28) (Fig. 3D). For both crop and noncroplands. the partial dependence between SOC and sample depth shows a negative association between depth and SOC, aligning with findings from previous literature. These relationships not only validate the effectiveness of the GAM developed for our study but also shed light on other critical questions regarding interactions between carbon dynamics in the soil and the atmosphere.

High acidic soils (low pH) often have slower decomposition rates, leading to higher SOC content, while highly alkaline soils (high pH) can promote decomposition and reduce SOC accumulation (43, 44), which is to some extent reflected in the results presented in Fig. 3C. While some studies suggest a strong positive correlation between precipitation and SOC (31, 45, 46), others demonstrate the minimal to no impact that precipitation has on SOC (34, 47). The ALEs calculated here (Fig. 3 E and F) reveal that at low to moderate precipitation SIs, the relation between SOC precipitation SI is complicated, while in highly humid or high vegetation environments, characterized by elevated SI or LAI, the presence of abundant vegetation does not necessarily result in increased SOC content [e.g., Spain (48) or Lu, Gilliam, Guo, Hou, and Kuang (44)]. This counterintuitive phenomenon can be attributed to specific factors and processes influencing carbon dynamics in these ecosystems. Accelerated erosion, high decomposition rates, rapid leaching of OM, the presence of high-quality and easily decomposable litter, prolonged waterlogging, and anaerobic conditions contribute to an accelerated net loss of SOC despite the vegetation cover.

Our findings indicate that the role of soil texture in influencing SOC stocks is complex and varies with different fractions of the soil (Fig. 3B). Specifically, we observed that increase in soil sand content is positively correlated with SOC content, especially in the 0 to 20% and 60 to 100% ranges. Similar trends are observed in other studies such as Vos et al. (49) and Yang et al. (50). At low SOC levels, clay and silt particles predominantly hold the SOC. However, as SOC content increases, the proportion of SOC in the sand fraction also rises (50). This suggests that only soils with a high sand fraction have the ability to retain high levels of SOC. Diverse contrasting positive and negative correlations between soil texture and SOC content have been reported (47, 51–54). The range of these relationships highlights the significance of SOC composition (55, 56) and clay type (57) as critical factors in the interplay between SOC and soil texture, especially at lower SOC levels.

For dominant land cover types (shown on the Insets, Fig. 4), we employed the fitted GAMs to estimate the topsoil OC (0 ~ 7 cm) content in soil sample locations as a result of a 1 SD increase in soil salinity (Fig. 4). For croplands, the 1 SD increase in soil salinity corresponds to an approximately 263.2% increase in the median salinity (relative to original median = 4.32 dS m⁻¹), while for noncroplands, it represents a nearly 245.24% increase in the median (original median = 3.83 dS m⁻¹). The predictions, along with 95% CIs, were made both before and after the increase in salinity levels. Before the 1 SD increase in soil salinity, the 95% CIs for the predictions showed a median range of 12.48 and 22.78 g kg⁻¹, for crop and noncroplands, respectively. However, after the increase in these parameters, the median 95% CIs were, respectively, 12.78 g kg⁻¹ for soil salinity of croplands (SI Appendix, Fig. S2) and 21.6 g kg⁻¹ for noncroplands (SI Appendix, Fig. S3). The results indicated higher uncertainty in the estimates for soil samples located in latitudes above 55°N, particularly in regions like Scandinavia and the northern United Kingdom.

For croplands, fitted GAM estimated a 1 SD increase in soil salinity is correlated with a decrease of approximately ~4.4% in the mean topsoil OC content of the soil samples (Fig. 4A), relative to the current mean predicted ${\overline{\mathit{SOC}}}_{\mathit{current}}$ of ~17.87 g kg⁻¹. On the other hand, a 1 SD increase in soil salinity of samples located in noncroplands was estimated to be correlated with an increase of approximately ~9.26% in the mean topsoil OC content (Fig. 4B; ${\overline{\mathit{SOC}}}_{\mathit{current}}$ = ~36.01 g kg⁻¹), in line with Setia et al. (13), who also estimated that world soils may lose 6.8 Pg of SOC by the year 2100 due to the projected increase in saline soils. To evaluate this, they used a modified Rothamsted Carbon model (58). Our results also indicated the variability in the relationship between soil salinity and SOC at different land covers. For croplands, the decrease in SOC in association with soil salinity was found to be little more noticeable (mean = ~−5.75%; ${\overline{\mathit{SOC}}}_{\mathit{current}}$ = ~23.71 g kg⁻¹) in lands covered by a mosaic of natural vegetation and cropland (tree, shrub, herbaceous cover >50% while cropland <50%) and less in irrigated croplands (mean = ~−3.78%; ${\overline{\mathit{SOC}}}_{\mathit{current}}$ = ~13.16 g kg⁻¹). However, for the noncroplands, the largest association was estimated for evergreen needle-leaved tree cover (mean = ~−14.25%; ${\overline{\mathit{SOC}}}_{\mathit{current}}$ = ~44.4 g kg⁻¹) and the lowest was estimated for grasslands (mean = ~−4.58%; ${\overline{\mathit{SOC}}}_{\mathit{current}}$ = ~33.57 g kg⁻¹).

In conclusion, this study provides insights into the complex interactions between soil salinity and SOC content across various land covers. The findings indicate that environmental factors that significantly explain SOC variability varying across different geographical regions and land use types. These results contribute to our understanding of the dynamics of SOC storage and have implications for land management strategies to mitigate the effects of changing environmental conditions on SOC levels. Analyzing the relationship between SOC content and soil salinity provides policymakers with insights to make informed decisions regarding sustainable land management, climate change mitigation, and agricultural practices. By integrating these findings into policies and management strategies, policymaking can use the results to promote soil health, enhance carbon sequestration, and contribute to global efforts to combat climate change.

We began by gathering soil profile data from soil databases, including measurements of SOC and soil salinity. Additionally, we acquired a complete dataset of various environmental and soil physio-chemical properties that are known to correlate with the SOC pool over medium- to long-term periods.

Following the preparation of the dataset, we selected and used the most significant parameters as an input for developing GAMs. Using the GAMs, we could recognize the specific correlation between soil salinity and SOC content while accounting for the effects of other environmental and physio-chemical factors.

We gratefully acknowledge The Climate and Environmental Research Institute NILU for providing the computational and storage resources and facilities essential for conducting this research (project NILU #B106057). Additional funding through the project SIS-EO (NILU #B121004) is gratefully acknowledged. N.S. would like to acknowledge funding for AI4SoilHealth project from the European Union’s Horizon Europe research and innovation programme under grant agreement No.101086179.