The Role of Soil Salinization in Shaping the Spatio-Temporal Patterns of Soil Organic Carbon Stock

Abstract

Soil salinization is closely related to land degradation, and it is supposed to exert a significant negative effect on soil organic carbon (SOC) stock dynamics. This effect and its mechanism have been examined at site and transect scales in previous studies while over a large spatial extent, the salinity-induced changes in SOC stock over space and time have been less quantified, especially by machine learning and remote sensing techniques. The main focus of this study is to answer the following question: to what extent can soil salinity exert an additional effect on SOC stock over time at a larger spatial scale? Thus, we employed the extreme gradient boosting models (XGBoost) combined with field site-level measurements from 433 sites and 41 static and time-varying environmental covariates to construct methods capable of quantifying the salinity-induced SOC changes in a typical inland river basin of China between the 1990s and 2020s. Results showed that the XGBoost models performed well in predicting the soil electrical conductivity (EC) and SOC stock at 0–20 cm, with the R² value reaching 0.85 and 0.81, respectively. SOC stock was found to vary significantly with increasing soil salinity following an exponential decay function (R² = 0.27), and salinity sensitivity analysis showed that soils in oasis were expected to experience the largest carbon loss (−137.78 g m⁻²), which was about 4.84, 14.37, and 25.95 times higher than that in the saline, bare, and sandy land, respectively, if the soil salinity increased by 100%. In addition, the decrease in the soil salinity (−0.32 dS m⁻¹) from the 1990s to the 2020s was estimated to enhance the SOC stock by 0.015 kg m⁻², which contributed an additional 10% increase to the total SOC stock enhancement. Overall, the proposed methods can be applied for quantification of the direction and size of the salinity effect on SOC stock changes in other salt-affected regions. Our results also suggest that the role of soil salinization should not be neglected in SOC changes projection, and soil salinization control measures should be further taken into practice to enhance soil carbon sequestration in arid inland river basins.

1. Introduction

Soils store about 1500 Pg (1 Pg = 10¹⁵ g) organic carbon in the first meter, which is almost double that in the atmosphere and three times that in the above-ground biomass [1]. As a fundamental component of the terrestrial ecosystem carbon cycle, soil organic carbon (SOC) plays an important role in the global carbon budget and exhibits a marked response to climate warming and anthropogenic activities [2,3,4]. In addition, SOC is closely linked to most ecosystem services, including soil structure and fertility improvement, food and fiber supply, soil and water retention, and cultural and scientific benefits [5,6,7]. Thus, monitoring and modeling of the spatial distribution and dynamics of SOC stock are currently a global focus attracting attention from researchers, policy makers, and socioeconomists for the purpose of carbon neutrality, food security, and carbon trading in the next decades [8,9,10,11]. SOC stock is essentially determined by the long-term balance between carbon inputs and mineralization as affected by vegetative, pedological, and biochemical processes, which are closely related to climate change and anthropogenic activities [3,7]. Of these biotic and abiotic influencing factors, soil salinization is closely related to land degradation and supposed to exert a significant effect on SOC dynamics [12,13]. However, our understanding of the role of soil salinization in shaping the spatio-temporal patterns of SOC stock in a broad scale (>10⁴ km²) remains limited.

Saline soils cover an area of 3.97 × 10⁶ km², accounting for about 3.10% of the total land area of the world [12]. Soil salinization is regarded as a predominant land degradation threat, which exerts a significant impact on the provisioning, regulating, supporting, and cultural services of ecosystems, leading to a reduction in soil fertility, production, stability, and biodiversity [14,15,16]. Thus, many soil salinization management practices have been taken into practice in salt-affected soil all over the world to enhance the ecosystem services of both croplands and degraded natural ecosystems. Generally, soil salinization can be divided into “primary” and “secondary” salinization according to the difference in the driving factors [17]. The main origins of salts in “primary” salinization are generally from precipitation, parent materials weathering, and the transport of saline dust while “secondary” salinization is mainly caused by intense agricultural activities [18,19,20]. The soil salinization processes are characterized by high dynamism [21,22]; thus, to fully evaluate the role of soil salinity in shaping the spatio-temporal patterns of SOC stock at a broad scale, updated spatial and temporal information of soil salinity with high resolution and precision should first be obtained.

Numerous researchers have proved that the combination of site-level soil measurements, remote sensing data, and machine learning models can enable cost-effective mapping of the spatio-temporal distribution of soil salinity from local to global scales [15,22,23,24,25,26]. The advantages of remote sensing data make it possible to directly observe the dynamics of soil surface salinity at various scales and resolutions [26,27]. The close relationships between the soil salt content and land surface reflectance, especially in the visible and near-infrared bands, make remote sensing monitoring capable of capturing the spatio-temporal variability of soil salinity [28]. In addition, salt indexes derived from spectral bands can also improve the prediction accuracy for soil salinity mapping. The common methods used to detect the relationships between remote sensing-based indicators and soil salinity include statistical regression models and machine learning. However, most statistical regression models are characterized by large errors in predicting higher or lower values for regions characterized by large spatial variability. In contrast, more accurate estimations can be made by machine learning models, due to their superior capability in detecting the nonlinear relationships between soil salinity and environmental covariates [26,29,30].

To determine the role of soil salinity in affecting SOC stock, the calibrated process-based soil carbon models, by accounting for the influence of salinity on the plant and decomposition rate, can be applied to estimate the possible SOC stock changes over space and time. Among the SOC models, RothC, CENTURY, and DNDC are widely used for carbon accounting [2]. Setia et al. [12] simulated the historic loss of SOC due to soil salinization using the RothC model and estimated that soils have experienced an average loss of 3.47 t ha⁻¹ due to salinization across the globe. The data-driven models based on machine learning models were also used for modeling SOC stock changes over space and time from regional to continental scales [7,11,31,32,33,34]. For instance, Li et al. [4] quantified the effects of climate warming and precipitation variation on SOC stock and projected its future trends based on a random forest model, and SOC stocks in the top 100 cm were projected to decrease significantly due to more carbon release as a result of the stimulated decomposition rates as affected by climate warming and deepened active layers in the permafrost. Similarly, Gautam et al. [35] employed ensemble machine learning models to estimate the baseline and dynamics of topsoil carbon stocks at a spatial resolution of 100 m in the United States and showed a total loss of 1.80 Pg carbon under the high-emission scenario by 2100. However, previous studies using the data-driven models for SOC stock prediction mainly focused on the effects of climate and land-use change, and until recently, the ability of machine learning-based data-driven models to detect the salinity-induced SOC dynamics has not been addressed, especially in inland regions [36,37].

In inland river basins, salts accumulation in soils is particularly intensive due to the strong potential evapotranspiration, excessive groundwater exploitation, and endorheic drainage, resulting in severe soil and vegetation degradation [27,38,39]. Over the past decades, a series of measures have been taken into practice for land restoration and reclamation in salinity-affected soils, and the changed salinity conditions may exert additional effects on SOC stock changes. Although the effects of soil salinization on the stock and cycling of SOC and the potential mechanism have been examined at a site or transect scale in previous studies [13,40,41], the effects of soil salinization control on SOC over a large spatial extent have been less quantified, especially by machine learning and remote sensing-derived soil salinity information. In addition, how to separate the effect of soil salinization on SOC stock from that of other environmental controls has been studied even less. Thus, to determine the role of soil salinity in shaping the spatio-temporal patterns of SOC stock, we selected a typical inland river basin of China as the study area, which is characterized by large areas of salt-affected soils, intensive ecological restoration, and agricultural activities. The main objectives of this study were to: (1) obtain soil salinity and SOC stock information with high spatial resolution and precision from the 1990s to the 2020s in the study area, and to (2) to separate the effect of soil salinization on SOC stock changes from that of other environmental controls over a large spatial extent.

2. Materials and Methods

2.1. Study Area

This study was conducted at the lower reaches of the Shiyang River Basin (38.0°–39.5°N, 101.8°–104.2°E), a typical arid inland river basin of China (Figure 1A). The study area covers an area of 1.62 × 10⁴ km², with elevation ranging from 1330 to 1703 m a.s.l. The region is characterized by a temperate desert climate, with the mean annual temperature and precipitation ranging from 6.50 to 8.20 °C and 99.8 to 124.7 mm, respectively. The dominant plant species include Nitraria tangutorum, Hedysarum scoparium, Salcola passerine, and Haloxylon ammodendron in desert regions; Tamarix chinensis, Lycium ruthenicum, Caragana korshinskii, and Artemisia ordosica in desert-oasis ecotone; and Zea mays, Triticum aestivum, and Helianthus annuus in oasis. The main land-use types are sandy land, bare land, oasis, and saline land, which account for 38.77%, 37.82%, 11.87%, and 10.69% of the total land area, respectively. Accordingly, the main soil types are classified as Haplic Arenosols, Haplic Gypsisols, Haplic Fluvisols (Arenic), and Haplic Solonchaks according to the World Reference Base for Soil Resources [42], which are roughly referred to as Aeolian Soils, Brown Desert Soils, Irrigated Desert Soils, and Solonchaks according to the Chinese Soil Genetic Classification [43]. The saline soils are generally characterized by low soil organic matter (<1%), with an average pH of 8.33. The main ions include Cl⁻, Na⁺, and SO₄²⁻, accounting for about 94% of the total ions amount in saline soils.

The study area is adjacent to the Tengger Desert and Badain Jaran Desert (Figure 1A); thus, it serves as an important ecological barrier in preventing the closure of the two big deserts in the Hexi Region [44,45]. However, with the rapid growth of the population and increasing grain demand, a large area of saline land and sandy land was cultivated to create an artificial oasis. Due to limited precipitation and surface runoff, agricultural irrigation in the study area mainly depended on groundwater exploitation, which was further strengthened by intense water use in the upper and middle reaches over the past decades. As a result, the basin had experienced severe environmental risks, such as a decrease in groundwater table depth, soil salinization, and land desertification, and degradation of groundwater-dependent desert ecosystems [46]. To prevent the ecological degradation trend, the Shiyang River Basin Comprehensive Management has been implemented since 2006, and a series of water resources management measures and ecological restoration projects were conducted after that. Considering the large areas of saline lands and frequent transfer of land degradation and development under intense agricultural activities and ecological restoration, the basin was supposed to be an ideal test region to evaluate the role of soil salinity in affecting SOC stock over space and time.

2.2. Methods

In this study, we employed the digital soil mapping framework to quantify the spatio-temporal patterns of soil EC and SOC stock for the 1990s, 2000s, 2010s, and 2020s based on machine learning models combined with 41 environmental covariates derived from remote sensing images, climate, topography, and land-use change data, and further determined the salinity-induced SOC changes based on sensitivity and scenario analysis (Figure 2). The basic theory of the framework is the SCORPAN equation, which expresses soil characteristics as functions of soil-forming factors such as soil inner properties, climate, organisms, terrain attributes, parent materials, time, and locations [47,48]. The spatio-temporal patterns of target soil properties (soil salinity and SOC stock in this study) can be inferred or predicted over space and time based on the trained relationships between soil properties and soil-forming factors and the spatial information of these environmental covariates [49]. In this study, we used machine learning models to determine the relationships between target soil properties and environmental covariates. The main steps involved (1) the collection of site-level measurements of soil salinity and SOC stock for the training of machine learning models, (2) processing and overlaying environmental covariates in raster format and linking them with site-level measurements, (3) detecting the relationships between site-level soil data and environmental covariates based on supervised machine learning models followed by cross-validation to evaluate and optimize the performance of models, and (4) inputting the gridded environmental covariates into the best-fitted machine learning models and obtaining the spatio-temporal patterns of soil EC and SOC stock.

2.2.1. Site-Level Soil Measurements

During the growing season (from May to October) of 2020 and 2021, in total, we investigated 197 sites mainly located in saline, sandy, and bare land along the salinity gradients across the study area (Figure 1B), including none-saline, slightly-saline, moderately-saline, strongly-saline, and extremely-saline lands and salt lake (Figure 1C). In each site, three 10 m × 10 m plots were selected to determine the species composition and vegetation cover, and geographic and topographic information such as longitude, latitude, elevation, slope, and aspect was also recorded using a GPS and a compass. Following this, we used a 35-mm-diameter soil auger to collect soil samples in 0–20 cm at 9 randomly selected sampling points for each plot. We mixed the 27 samples from the 3 plots to generate a composite soil sample for each sampling site. In addition, we determined the soil bulk density using a soil auger equipped with a stainless-steel cylinder (5.5 cm in diameter and 4.2 cm in height) to obtain intact soils in each of the three plots. The SOC content was measured by the wet Walkley–Black method [50], and the bulk density was calculated from the soil cores as the ratio of the oven-dry soil mass (after 24-h desiccation at 105 °C) to the core volume. The soil electrical conductivity (EC_1:5) of the samples was determined using the soil leachate at a 1:5 soil-to-water ratio under a 25 °C environment temperature. The background value of soil EC_1:5 for these samples reached 1.95 dS m⁻¹, indicating severe soil salinization in the study area. We also obtained SOC content, bulk density, and soil EC_1:5 data of an additional 236 sites from Chen et al. [51] and Li et al. [52], which were mainly located in salinity-affected oasis and oasis-desert ecotone. The total site-level soil EC and SOC measurements used in this study were 433 and 313, as some previous sites were lacking SOC measurements. We calculated SOC stock at 0–20 cm according to the following equation:

$$\mathrm{SOCS} = \mathrm{BD} \times \mathrm{D} \times \mathrm{SOC} \times (1 - \mathrm{G}) \times 0.01$$

(1)

where SOCS is SOC stock (kg m⁻²); BD, D, and SOC represent the bulk density (g cm^–3), soil depth (cm), and SOC content (g kg^–1); and G is the volumetric percentage occupied by gravel > 2 mm (%), which was set at zero due to trivial gravel contents in saline, sandy, and oasis soils in the study area.

2.2.2. Environmental Covariates

In total, we used 41 environmental covariates to train the machine learning models, compromising remote sensing images (vegetation, salt, and brightness indexes), terrain attributes, climate, land use and land cover, and geographic locations (Table S1 [53,54]). In a broad sense, these covariates could be divided into two groups, i.e., time-varying and static. The time-varying covariates included remote sensing-based indexes, climate, and land use and land cover, which could account for the decadal changes in soil salinization and carbon accumulation. Climate factors, including temperature and precipitation, were derived from the high-spatial-resolution monthly temperature and precipitation datasets, which were generated by spatially downscaling the Climatic Research Unit (CRU) TS v4.02 long-term climate datasets based on WorldClim 2.0 30″ (about 1 km at the equator) datasets [54]. As the climate influenced the SOC stock less on a yearly basis, a 5-year averaging window was applied to process the climate data. The land use and land cover data with a 30 m resolution for the 1990s, 2000s, 2010s, and 2020s was acquired from the Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences (http://www.resdc.cn (accessed on 6 January 2022)). This land-use product, known as the National Land-Use/Cover Database of China (NLUD-C), was primarily produced based on Landsat TM and OLI images by visual interpretation, with a greater emphasis on geometric correction. This product contained 29 land-use types in total with a classification accuracy of more than 90% [55]. In this study, we reclassified the raw category into 6 main types, i.e., the sandy land, saline land, oasis, bare land, build-up land, and water, for follow-up analysis. In addition, we obtained the water amount and water depth data from the Shiyang River Administration.

The remote sensing-based vegetation, salt, and brightness indexes used in this study were derived from the Landsat TM, ETM+ and OLI images. We collected Landsat images with cloud cover less than 20% during the growing season of 1990–2020 based on the Google Earth Engine (https://code.earthengine.google.com/ (accessed on 6 July 2021)). Radiometric calibration, atmospheric correction, cloud removal, and clipping were first conducted for the images. In addition, as some of the images were characterized by low quality, especially for the ETM+, a 5-year window was used to make a composite image with maximum NDVI values using the qualityMosaic function. The blue, green, red, and near-infrared bands were employed as spectral indices to further calculate other remote sensing indexes (Table S1). The efficiency of these indexes in predicting soil EC or SOC stock was confirmed by previous studies [26,27,29,30]. It should be noted that the sensors and band width differences among Landsat 5, 7, and 8 may cause uncertainty regarding the consistency of temporal predictions [56]. To minimize this uncertainty, we harmonized the reflectance values of each band before obtaining the composite images for each period, based on linear regression methods with more than 2000 calibration sites covering all main land cover types in the study area (Figures S1 and S2). The operation period of ETM+ overlapped with TM and OLI over time; thus, we could use ETM+ as a link to determine the regression coefficient of the three sensors (Table S2).

Static covariates were mainly terrain attributes and geographic locations, which were assumed to remain constant through the study period. Terrain attributes, including elevation, slope, aspect, topographic wetness index, multi-scale topographic position index, and potential radiation, were all derived from the 30-m digital elevation model (DEM) derived from the ASTER GDEM, based on the SAGA GIS software (https://search.earthdata.nasa.gov/search/ (accessed on 10 January 2022)). All environmental covariates used for digital soil mapping were resampled to 30 m by the bilinear interpolation (nearest-neighbor interpolation for categorical variables), and further georeferenced to the WGS 1984 coordinates system in ArcGIS 10.2 (ESRI Inc., RedLands, CA, USA). The values of the static environmental covariates at each sampling site were directly extracted from the raster values except the elevation, slope, and aspect, which were already precisely recorded during field sampling. For the dynamic environmental covariates, we extracted values at each sampling site corresponding to that particular year.

2.2.3. The XGBoost Model

In this study, we first trained the machine learning model based on the final prepared matrices with 42 columns (41 representing covariates and 1 for the target soil property) and 433 rows. As for the specific training process, we employed an extreme gradient boosting (XGBoost) model to detect the nonlinear relationships between soil EC and environmental covariates. The XGBoost model is a popular and efficient open-source implementation of the gradient boosted trees algorithm developed by Chen and Guestrin (2016) [57]; it runs much faster, and uses far fewer computation resources and internal storage than other classic machine learning models. The gradient boosting is a supervised learning algorithm that aims to accurately predict a target variable by combining the estimates of a set of simpler and weaker models, i.e., “boosting”, as a basic idea of this method extends a “strong” learner from a set of “weak” learners [57]. When employing the XGBoost model for regression, the weak learners are actually regression trees, and each regression tree generates an input data point to one of its leaves that contains a continuous score. The XGBoost model minimizes a regularized objective function that combines a convex loss function (based on the difference between the predicted and target outputs) and a penalty term for model complexity (i.e., the regression tree functions) (Figure 3). The training proceeds iteratively, adding new trees that predict the residuals or errors of prior trees that are then combined with previous trees to make the final prediction. The model is called gradient boosting because it uses a gradient descent algorithm to minimize the loss when adding new models. Parallel and distributed computing makes learning faster, which enables quicker model exploration. Thus, the XGBoost model is quite suitable for high-resolution soil mapping [58,59].

2.2.4. Soil EC and SOC Stock Prediction

We conducted an environmental covariates selection procedure before model training. Figure S3 shows the Pearson correlation coefficients among all covariates used in the modeling and prediction of soil EC and SOC stock, and we found that most of them are self-correlated. To ensure the computational efficiency, we only used the top 20 covariates according to the relative importance values (Figure S4). We used 10-fold cross-validation procedures to determine the optimal values for the parameters in the model [60]. We finally set the maximum depth of a tree (max_depth), learning rate (eta), and max number of boosting iterations (nround) at 6, 0.45, and 50, respectively. For SOC stock modeling, we used soil EC as an additional predictor. The performance of the optimal machine learning model in predicting soil EC and SOC stock was indicated by statistics, including the mean absolute error (MAE), root-mean-square-error (RMSE), coefficient of determination (R²), and Lin’s concordance correlation coefficient (LCCC), based on 30 runs of the 10-fold cross-validation procedure [60].

The soil EC and SOC stock maps for the 1990s, 2000s, 2010s, and 2020s were obtained based on the trained models with spatially gridded environmental covariates in the corresponding period as inputs. It should be noted that we predicted the soil EC and SOC stock at a 30 m resolution and mapped both the means and uncertainties. We had 8801 × 5165 pixels for each environmental covariate; thus, the random-access memory (RAM) and computing speed demands were quite large to produce high-resolution soil EC and SOC stock maps. Thus, the final raster layers were stacked and then split into 59 tiles of 20 km × 20 km for calculation. We converted each tile into a ‘data.table’ object containing x-y (representing longitude and latitude) and values of predictors for each pixel, and further transformed the ‘data.table’ to ‘xgb.Matrix’ to perform predictions with the trained XGBoost models under the R language environment [61]. The predictions and x-y coordinates in the output tables were converted into raster format for each tile, and the final maps were generated by merging all 59 tiles.

We applied a non-parametric bootstrap method to quantify the uncertainties in XGBoost modeling by bootstrapping the training matrices 30 times and generating 30 XGBoost models [59]. The uncertainties were expressed as standard deviations of the 30 bootstrap predictions. The “xgboost”, “Matrix”, and “data.table” packages were employed to conduct the XGBoost algorithm, and the “getSpatialTiles” function in the “landmap” package was employed for building tilling systems, and the “raster”, “rgdal”, and “sp” packages were also used for processing predictors in raster formats. Parallel computation was conducted to improve the calculation efficiency based on the “snowfall” and “parallel” packages under the R language environment.

2.2.5. Quantification of the Salinity Effect on SOC

To separate the effect of salinity on SOC stock from that of other environmental controls such as vegetation and climate, we conducted a sensitivity experiment by only altering the values of soil EC while keeping other environmental covariates constant. A salinity sensitivity analysis for SOC stock could enable us to quantitatively evaluate the size and direction of SOC stock change in response to soil salinization over space and time. We applied SOC stock, soil EC, and other environmental covariates data of the 2010s to define baseline conditions, and another 10 soil EC change scenarios (from −100%, −80%, −40%, −20%, −10%, +10%, +20%, +40%, +80% to +100%) were set for SOC stock prediction.

In addition, we set an additional scenario, i.e., the salinity-stable scenario, which assumed that soil salinity has been constant over time since the 1990s while other environmental covariates have varied with time. Specifically, we obtained two SOC stock maps for each of the three periods (2000s, 2010s, and 2020s), of which one was predicted based on predictors in the corresponding period while the other was predicted based on the same predictors except soil EC being substituted by values from the 1990s. We then calculated the SOC stock difference between the two SOC stock maps.

We finally obtained the spatial distribution of the SOC stock differences (δSOC stock) between each scenario and baseline and summarized the means and standard deviations for each land-use type. It should be noted that we excluded other remote sensing-based salinity indexes during the SOC stock prediction for each scenario. This was mainly because we already had included soil EC in the model as a predictor; thus, the introduction of other salinity indexes into the models may have masked the effects of salinity on SOC stock changes.

3. Results

3.1. Soil Salinity and SOC Stock Modeling

The site-level soil EC at 0–20 cm ranged from 0.01 to 15.27 dS m⁻¹ with an average of 1.95 dS m⁻¹ over the study area (Figure 4). The mean SOC stock at 0–20 cm was 1.23 kg m^–2 and varied from 0.01 to 3.88 kg m^–2. The coefficients of variation for SOC stock and EC were high (>80%) and showed high variability across the study area. The XGBoost models could simulate the non-linear relationships between SOC stock, soil EC, and other environmental covariates well, with a slope of 0.87 and an intercept of 0.20 dS m^–1 for soil EC (Figure 4A) and 0.87 and 0.16 kg m^–2 for SOC stock (Figure 4B), based on the linear fits between the predicted and measured values from the validation datasets (Figure 5). The results of the 10-fold cross-validation also showed that the R² and RMSE values of the models were 0.85 and 1.41 dS m^–1 for soil EC and 0.81 and 0.46 kg m^–2 for SOC stock, respectively, indicating ideal performances of the XGBoost models in predicting the spatio-temporal patterns of soil salinity and SOC stock (

Table 1. Performance of the XGBoost models in predicting soil EC and SOC stock.

Variable	Statistics	Minimum	1st Quartile	Median	Mean	3rd Quartile	Maximum	Standard Deviation
Soil EC (dS m−1)	MAE	0.57	0.58	0.60	0.60	0.61	0.62	0.02
	RMSE	1.28	1.37	1.41	1.41	1.45	1.51	0.06
	R2	0.81	0.84	0.85	0.85	0.85	0.88	0.02
	LCCC	0.89	0.91	0.91	0.91	0.92	0.93	0.01
SOC stock (kg m−2)	MAE	0.30	0.31	0.32	0.32	0.32	0.33	0.01
	RMSE	0.45	0.45	0.46	0.46	0.47	0.49	0.01
	R2	0.79	0.80	0.81	0.81	0.81	0.82	0.01
	LCCC	0.89	0.90	0.90	0.90	0.90	0.90	0.00

).

3.2. Spatio-Temporal Patterns of Soil Salinity and SOC Stock

The area-weighted mean soil EC at 0−20 cm over the study area decreased from 1.17 to 0.85 dS m⁻¹ while the SOC stock increased from 0.35 to 0.51 kg m⁻² during the 1990s−2020s (

Table 2. Summary of soil EC and SOC stock by land use and land cover type.

Variable	Type	Periods
		1990s	2000s	2010s	2020s
Soil EC (dS m–1)	Bare land	0.63 ± 0.13	0.65 ± 0.12	0.61 ± 0.11	0.48 ± 0.28
	Built-up land	1.19 ± 0.14	1.14 ± 0.16	1.12 ± 0.15	1.29 ± 0.48
	Oasis	1.18 ± 0.14	1.12 ± 0.16	1.03 ± 0.13	0.96 ± 0.24
	Saline land	4.78 ± 0.30	4.37 ± 0.30	4.52 ± 0.31	3.38 ± 0.91
	Sandy land	0.54 ± 0.09	0.59 ± 0.10	0.58 ± 0.09	0.49 ± 0.34
	Area-weighted mean	1.17 ± 0.14	1.13 ± 0.14	1.07 ± 0.13	0.85 ± 0.12
SOC stock (kg m–2)	Bare land	0.29 ± 0.06	0.35 ± 0.07	0.40 ± 0.07	0.37 ± 0.08
	Built-up land	1.28 ± 0.12	1.33 ± 0.14	1.49 ± 0.16	1.40 ± 0.20
	Oasis	1.28 ± 0.15	1.54 ± 0.15	1.69 ± 0.18	1.88 ± 0.21
	Saline land	0.38 ± 0.06	0.43 ± 0.08	0.47 ± 0.08	0.51 ± 0.09
	Sandy land	0.17 ± 0.04	0.19 ± 0.05	0.23 ± 0.05	0.22 ± 0.05
	Area-weighted mean	0.35 ± 0.03	0.41 ± 0.03	0.49 ± 0.03	0.51 ± 0.03

). Spatially, over the past three decades, both the means and standard deviations of soil EC were estimated to decrease in the northern and southeastern parts of the study area, especially after the 2010s (Figure 6). By contrast, SOC stock in the northern and southeastern parts was estimated to experience a rapid increase throughout the study period (Figure 7). Temporally, soil EC experienced a greater decrease after the 2000s, especially during the 1990s−2020s, mainly contributed by saline land and oasis (Table 2). However, SOC stock was estimated to experience the maximal (0.08 kg m⁻²) and minimum increase (0.02 kg m⁻²) during the 2000s−2010s and 2010s−2020s, respectively, with the largest increase occurring in oasis. The uncertainties of soil EC and SOC stock showed similar spatio-temporal patterns in comparison with the means (Figures S5 and S6).

3.3. Effects of Salinity on SOC Stock

We found that SOC sock decreased significantly with increasing soil EC, and their relationship could be expressed as an exponential decay function, i.e., y = 0.23 + 1.62exp(−0.28x), with the R² value reaching 0.27 (Figure 8). We also found that the decreasing trend was mainly contributed by data points from oasis, and SOC stock in bare land was less varied with soil EC. In addition, SOC stock in sandy lands seemed to exhibit a slightly increasing trend with increasing soil EC. Salinity sensitivity analysis showed that the responses of SOC stock to salinity variations exhibited strong spatial heterogeneity. A reduction in soil EC led to an SOC stock increase in most parts of the study area, and the higher the decrease in soil EC, the larger the values for SOC stock (Figure 9, Figure 10 and Figure 11). In contrast, the increase in soil EC will significantly decrease SOC stock. The area-weighted mean SOC stock was estimated to change by 0.040 and −0.025 kg m⁻² if soil salinity varied by −100% and 100%, respectively, with marked changes occurring in oasis areas (

Table 3. Summary of the area-weighted averages and standard deviations of SOC stock differences (δSOC stock) by land use and land cover type under different salinity change scenarios.

Soil EC Change	δSOC Stock (g m−2)
	Bare Land	Build-Up Land	Oasis	Saline Land	Sandy Land
−100%	23.20 ± 11.35	113.99 ± 44.95	139.83 ± 48.02	71.52 ± 19.11	17.89 ± 8.13
−80%	15.84 ± 10.17	96.21 ± 42.32	116.88 ± 44.86	48.06 ± 17.89	12.13 ± 6.92
−40%	6.69 ± 7.66	56.60 ± 37.74	67.49 ± 40.60	30.58 ± 13.50	4.89 ± 4.77
−20%	3.08 ± 5.47	27.02 ± 29.85	34.10 ± 32.50	13.69 ± 8.36	2.22 ± 2.98
−10%	1.48 ± 3.60	13.24 ± 20.26	16.76 ± 22.18	6.16 ± 5.17	1.05 ± 1.92
+10%	−1.33 ± 3.47	−12.74 ± 19.81	−16.24 ± 22.03	−4.66 ± 4.54	−0.89 ± 1.79
+20%	−2.55 ± 5.12	−25.77 ± 29.59	−32.47 ± 32.79	−8.71 ± 6.73	−1.69 ± 2.66
+40%	−4.72 ± 6.92	−52.96 ± 39.70	−65.28 ± 43.61	−15.07 ± 9.30	−2.98 ± 3.87
+80%	−8.22 ± 8.97	−95.54 ± 47.69	−120.39 ± 52.42	−24.68 ± 12.40	−4.70 ± 5.35
+100%	−9.59 ± 9.53	−107.82 ± 48.53	−137.78 ± 53.55	−28.44 ± 13.30	−5.31 ± 5.83

). In addition, a reduction in salinity over the past three decades was estimated to enhance SOC stock by 0.015 kg m⁻², and this amount was about 10% of the total SOC stock increase (Figure 12). δSOC stock in oasis (94.24 g m⁻²) exhibited a stronger response to salinity change than in saline (23.34 g m⁻²), sandy (6.03 g m⁻²), and bare land (−3.43 g m⁻²) (

Table 4. Summary of the area-weighted averages and standard deviations of SOC stock differences (δSOC stock) by land use and land cover type under salt-stable scenarios for the 2000s, 2010s, and 2020s.

Type	δSOC Stock (g m−2)
	2000s	2010s	2020s
Bare land	6.29 ± 18.47	−7.38 ± 22.29	−3.43 ± 24.11
Built-up land	53.85 ± 40.29	58.70 ± 45.02	47.20 ± 46.86
Oasis	76.32 ± 46.48	73.84 ± 50.11	94.24 ± 54.18
Saline land	8.77 ± 17.00	7.72 ± 18.63	23.34 ± 21.55
Sandy land	6.71 ± 14.21	1.37 ± 13.20	6.03 ± 12.99

).

4. Discussion

4.1. The Efficiency of the XGBoost Models in Predicting Soil Salinity and SOC Stock

Our results suggest that the XGBoost models behaved well in predicting the spatio-temporal patterns of soil EC and SOC stock in the study area. This is mainly because the XGBoost models are efficient tree boosting systems, which can better deal with the complex nonlinear relationships between the target soil properties and the environmental covariates in comparison with linear regression models [35,58,59]. In addition, the XGBoost models need not any statistical premise like independence and normality which were necessary in linear regression models, and the over-fitting issues are also controlled with a more regular model formalization method to produce more reliable predictions [57]. In this study, we used a large number of site-level measurements and 41 environmental covariates derived from the remote sensing, climatic, topographic, and land-use data to train the machine learning models. However, most of the predictors were self-correlated (Figure S3); thus, we finally used covariates with relatively high variable importance (Figure S4). We found that the XGBoost models allowed spatially explicit predictions of soil EC and SOC stock to be made, with the R² value reaching 0.85 and 0.81, and RMSE reaching 1.41 dS m⁻¹ and 0.46 kg m⁻², respectively, based on the 10-fold cross-validation procedures. Our results were close to that of previous studies conducted at a similar spatial scale, in which the machine learning models with dozens of covariates as model inputs were applied to mapping soil EC or SOC stock [4,7,26,33,62]. Considering the high efficiency and precision, the XGBoost models were recommended to predict the spatio-temporal patterns of soil salinity and SOC stock, especially at a higher resolution over a broad scale, using both the time-varying and static environmental covariates after variables’ selection.

4.2. Spatial Patterns of Soil Salinity and SOC Stock

We found that both soil EC and SOC stock varied significantly across the basin, being mainly affected by terrain attributes and vegetation, respectively (Figure S4). Specifically, soil salinity was mainly distributed in the northern and southwestern sub-regions, which were characterized by a lower elevation and higher topographic wetness index, confirming that “primary” salinization mainly occurred at tail-end-lakes in inland river basins. Actually, the terminal of the Shiyang River was a huge lake (~600 km²) called the Zhuye Lake in history, which has gradually shrunk because of reclamation and irrigation since the Han Dynasty [63]. Over the past century, increasing water resource demands in the upper and middle reaches of the Shiyang River due to intense agricultural activities have significantly lowered the water amount into the lower reaches, and construction of the Hongyasha desert reservoir since 1958 has further substantially reduced the area of the lake, which finally dried up in 1959. The shrinkage of terminal lakes led to the accumulation of salts in soils, which created large areas of salinity land across the study area.

The soil salinity in oasis was also quite high, with an area-weighted mean reaching to 0.96 dS m⁻¹ in the 2020s. According to the regression equation (y = 1.404 + 2.676x + 0.429x², Figure S7) between the soil salt content and soil EC_1:5 in the study area, the mean salt content in oasis reached 4.37 g kg⁻¹, i.e., moderately-saline lands according to the commonly used soil salinity classification scheme [64]. The high soil salt content in oasis was mainly associated with intense agricultural irrigation, which pumped large amounts of groundwater to the land surface and subsequently to air through evapotranspiration, leaving salts in the soils [39,65]. “Secondary” salinization occurred when the water table depth was 2–3 m from the land surface. The capillary action raised salts from groundwater to surface soils, especially in areas with high groundwater mineralization (the case in our study area). This process could be substantially accelerated by intense groundwater exploitation in extremely arid regions characterized by limited rainfall entering the aquifer, since almost all groundwater used for irrigation contained some dissolved salts [20,66]. In addition, soil salinization from irrigation water was also significantly strengthened by poor drainage and the use of saline water for irrigation [18,19,30], which was quite common in the lower reaches of the Shiyang River Basin [46].

As for SOC stock, it mainly varied with land use and land cover type across the study area. The oasis was characterized by the maximal SOC stock value reaching 1.88 kg m⁻² in the 2020s, followed by built-up land mainly located within the oasis. The oasis mainly compromised croplands, wetlands, and windbreak forests in the study area; thus, SOC stock was higher due to the higher root biomass input associated with greater water availability in comparison with that of zonal temperate desert ecosystems [45]. Being different from humid regions, cultivation in arid regions (especially in desert and bare land) generally enables soil to accumulate additional organic carbon instead of releasing carbon [67]. This is mainly because the standing organic carbon stock is very low in zonal soils (less than 2 g kg⁻¹ for topsoil) due to little plant carbon inputs under such arid environments. SOC stock could increase after cultivation as a result of increasing crops residues and root debris input into soils under sufficient water supply. The minimum SOC stock was observed in sandy land and bare land (mainly Gobi Desert), where the soils are characterized by very coarse texture, resulting in less mineral protection from fine particles for organic carbon [39]. SOC stock in saline land reached 0.51 kg m⁻², which was significantly higher than that in sandy and bare lands. This was because most saline land mainly originated from drying-up terminal lakes characterized by large areas of lake sediments; thus, the soil texture was finer than that of sandy land, leading to more old carbon being stored in fine mineral particles [3]. In addition, salt marsh also stored much higher carbon in soils than that of sandy land. Overall, the SOC stock in our study area was low, and only ecosystems with sufficient water supply could accumulate additional carbon in soils.

4.3. Temporal Dynamics of Salinity and SOC Stock

Soil EC in the lower reaches of the Shiyang River basin showed an overall decreasing trend while SOC stock increased significantly from the 1990s to the 2020s (Figure 6 and Figure 7, Table 2). In contrast, the results from Hassani et al. (2020) showed that the trends in the total area of soils with EC_e ≥ 4 dS·m⁻¹ between 1980 and 2018 were statistically meaningful (p < 0.05) for about 117 of 256 countries/states, among which the following had the highest rate of annual increase: Brazil, Peru, Sudan, Colombia, and Namibia, while for China, the area of salinity-affected soils exhibited a decreasing trend, which is consistent with the result of this study [66]. This was probably associated with the many ecological restoration engineering projects implemented in China over the past decades. In our study area, lots of ecological restoration practices were taken into practice due to the implementation of the Shiyang River Basin Comprehensive Management (SRBCM) in 2006 (Figure 13).

The study area has experienced significant climate warming over the past decades (Figure S8), and the natural ecosystems rapidly degraded due to stronger potential evapotranspiration and decreasing water table depth associated with groundwater overdraft (Figure 13B). To prevent the degradation trend of natural ecosystems, the SRBCM has been implemented since 2006. The total surface water amount into the lower reaches increased by 3.03 × 10⁸ m³ from 2005 to 2019 (Figure 13A). As a result, the issues of groundwater level decrease, water shortage, and ecological degradation were significantly alleviated, and the areas of oasis increased while the areas of sandy and saline lands significantly decreased, especially around the Qingtu Lake area, where the NDVI and water table depth have significantly increased since 2006 (Figure 13B, Figures S9 and S10). It was estimated that soil EC in saline lands slightly decreased from 4.78 to 4.37 dS m⁻¹ during the 1990s−2000s while it sharply decreased from 4.52 to 3.38 dS m⁻¹ during the 2010s–2020s. The implementation of the SRBCM in Qingtu Lakes removed salts from the soil surface layers due to vertical water washing processes, and cultivation and salinization management in saline land at the Nanhu town (southeastern part of the region) has also significantly reduced surface salt over the past decade (Figure 6 and Figure S9). Thus, the decreases in soil salinity in our study area over the past decades were mainly associated with ecological restoration practices.

The temporal dynamics of SOC stock in the study area was mainly associated with oasis expansion, vegetation restoration in sandy and saline regions, salinization control, and agricultural management (e.g., the application of organic and microbial fertilizers, farmyard manure, and crop residues), which together contributed to a 0.15 kg m⁻² increase in the topsoil carbon stock over the past three decades. In arid inland regions, water availability was a predominant control of ecosystem productivity; thus, it was also regarded as the main external environmental driver of SOC dynamics [45,60]. In contrast, vegetation restoration and salinization control were supposed to be regulatory factors, which could exert additional effects on SOC stock if the water availability was guaranteed for the arid ecosystems. The SRBCM has brought an additional 3.03 × 10⁸ m³ water resources into the study area since 2006, and this directly led to a significant conversion of saline and sandy land to oasis (Figure S9). The area of saline and sandy land decreased by 235.83 and 59.42 km², respectively, and oasis increased by 322.27 km² from the 2000s to the 2010s. In contrast, oasis only expended by 40.14 km², and the saline and sandy land shrank by 21.77 and 38.19 km², respectively, during the 1990s to 2000s (Table S3). As a result, the NDVI exhibited a significant increasing trend (R² = 0.64) in the study area over the past two decades. The overall improvement in vegetation conditions was supposed to contribute the largest proportion of the increase to SOC stock (Figure S4B). In addition, soil salinization control, organic fertility application, and tillage management such as straw/stover return also exerted positive effects on SOC accumulation in cultivated oasis, and these effects probably increased with each cultivation year due to higher plant carbon inputs and decreased soil pH in croplands than in the sandy and saline land [67]. Climate change also affected the SOC dynamics, and the significant increase in temperature and decrease in precipitation after the 2010s may be the reason that SOC decreased by 0.03 and 0.01 kg m⁻² in bare and sandy land during the 2010s−2020s (Figure S8 and Table 2).

4.4. Effects of Salinity on SOC Stock

The soil carbon–salinity relationship, based on site-level measurements, showed that SOC sock decreased significantly with increasing soil EC following an exponential function, with an R² value of 0.27 (Figure 8). Similar results were also obtained by previous studies conducted in other salt-affected regions [13,68]. Generally, SOC stock in saline soils is affected by two contrary processes: reduced plant inputs and lowered decomposition rates [12]. On the one hand, salinity stress directly reduces plant water uptake and deteriorates transpiration, resulting in a reduction in carbon inputs into soils. On the other hand, it also decreases microbial activities and the rates of organic matter decomposition, leading to an increase in SOC stock if the carbon inputs are unchanged [14,66]. In most cases, the negative effect of soil salinity on SOC stock is much larger than the positive effect, especially in agroecosystems [19,40]. In this study, we found that the negative relationship between SOC stock and soil EC was mainly contributed by data points from oasis while SOC stock in bare land varied less with soil EC. This result confirmed that salinity was a regulatory factor that exerted an additional effect on SOC stock only when the water availability was guaranteed for the arid ecosystems. In cultivated oasis, the vegetation-soil system was well irrigated, and salinity could exhibit an effect on carbon cycling by significantly altering the microbial activity and plant growth processes. In contrast, in the bare land of arid inland regions, the carbon cycling-associated biotic processes were trivial due to the extreme shortage of water resources, and the variation of the salinity could exert lesser effects on SOC stock. Surprisingly, we also found that SOC stock in sandy land seemed to exhibit a slightly increasing trend with increasing soil EC. We deduced that this was probably associated with the effect of soil texture. In sandy regions, soils with a greater salt content were generally characterized by finer soil particulates, which were originally derived from river and lake deposits containing more old carbon in comparison with coarse sand particles [14,39]. Nevertheless, the SOC stock in the study area showed a significantly negative relationship with soil salinity at the basin scale.

It should be noted that the carbon–salinity relationship obtained according to the spatially distributed site-level measurements was probably mixed with the vegetation and soil texture effects on SOC, as regions with high vegetation cover were generally characterized by lower soil salinity. To separate the effect of salinity from that of other environmental divers, we additionally conducted sensitivity and scenario analysis based on the XGBoost models. We found that the area-weighted mean SOC stock was estimated to increase by 0.040 kg m⁻² if the soil salinity only decreased by 100% (Table 3). The SOC stock dynamics also exhibited significant spatial heterogeneity, with the largest decrease occurring in oasis (−137.78 g m⁻²) if soil salinity increased by 100%. This change in oasis was about 4.84, 14.37, and 25.95 times higher than that in the saline, bare, and sandy land, respectively (Table 3). In contrast, the standing SOC stock in oasis was 1.69 kg m⁻², which amounted to 3.60-, 4.22-, and 7.35-fold greater than in the bare and sandy land, respectively (Table 2). Thus, SOC stock in the oasis was supposed to be the most sensitive to salinity variation. This conclusion again confirmed that salinity was a regulatory factor that could exert stronger effects on SOC stock if the water availability was guaranteed for the arid ecosystems. Our results also suggested that the effects of soil salinization should be considered in SOC modeling and future projection, especially in oasis of arid inland regions.

The salt-stable scenario analysis showed that the reduction in salinity over the past three decades was estimated to enhance SOC stock by 0.015 kg m⁻², which was about 10% of the total SOC stock increase as affected by oasis expansion and ecological restoration. Similar to the results of the salinity sensitivity analysis, the oasis (0.094 kg m⁻²) contributed to the largest proportion of SOC enhancement, followed by built-up land and saline land (Table 4 and Figure 12). This was mainly because the soil nutrients availability and vegetation growth were favored by decreasing salt contents as a result of salinity control over the study period. The reduction in salt stress stimulated plant productivity, which further enhanced the carbon input into soils. In addition, the changed microbial community composition also exerted a significant impact on carbon accumulation [41,69]. Changes in soil salinity reshaped the soil microbial community structure and activity, and the input of sulfate in the process of salinization further changed the main way of soil carbon metabolism [40,70,71]. Traditionally, enhanced soil microbial activity is supposed to accelerate the decomposition rate of carbon while soil microbes also tremendously contribute to carbon stabilization in croplands through the microbial carbon pump (MCP) [72]. The MCP suggests that soil microbes play an important role in carbon stabilization by continuously transforming labile carbon to persistent forms. For instance, Zhu et al. [73] showed that the necromass carbon derived from microbes contributed to the largest proportion of SOC accumulation in bio-energy croplands. In this study, the decrease in soil salinity could enhance plant productivity and microbial activity, which further enhanced carbon input into soils and the efficiency of MCP, leading to more stable microbial necromass carbon accumulation in soils. Our results highlight the potential of soil salinization control in additionally enhancing SOC stock in inland regions.

4.5. Uncertainties and Limitations

There are some uncertainties and limitations in our study. Although we used 41 time-varying and static environmental covariates (Table S1) to model the spatio-temporal dynamics of soil EC and SOC stock at a resolution of 30 m based on the digital soil mapping framework, most of the predictors were remote sensing-based. Other environmental drivers such as nitrogen deposition, climate extremes, CO₂ enrichment, and agricultural managements were not directly considered in modeling the SOC dynamics for this study [3,7,26,74,75]. This was mainly because these factors were hard to quantify in the machine learning models due to limited training datasets and spatially explicit products. In addition, in arid inland regions, biotic processes are predominantly affected by water, and other environmental controls may exert much less impacts on SOC stock in comparison to water availability. In this study, the effects of water availability on SOC dynamics could be captured by time-varying covariates, including precipitation, satellite-derived vegetation indexes, brightness index, and land use and land cover change data, over time (Table S1). Moreover, the main aim of this study was to quantify the role of soil salinity in shaping the spatio-temporal patterns of SOC stock. Thus, we did not fully consider the effects of CO₂ enrichment, nitrogen decomposition, and climate extremes in our modeling and results analysis. Nevertheless, we believe that a comprehensive accounting of these factors in the machine learning models will lower the uncertainties in SOC prediction over space and time and follow-up future studies should be conducted based on systematic sampling and spatially explicit gridded covariates.

5. Conclusions

Quantifying the saline control of SOC stock is important for developing targeted soil carbon management strategies in salt-affected regions. In this study, we employed the XGBoost models combined with remote sensing data to predict soil EC and SOC stock dynamics from the 1990s to the 2020s and separated the effects of soil salinity on SOC from that of other environmental drivers based on scenarios analysis. We found that the combination of sufficient site-level measurements with the XGBoost models allowed spatially explicit estimates of salinity-induced SOC stock changes to be made. Specifically, the area-weighted mean soil EC at 0−20 cm over the past three decades was estimated to decrease by 0.22 dS m⁻¹ while the SOC stock increased by 0.16 kg m⁻², mainly due to oasis expansion and ecological restoration. The SOC stock was found to vary significantly with soil salinity following an exponential decay function (R² = 0.27) at the basin scale. The salt sensitivity analysis further showed that if soil salinity decreased by 100%, the area-weighted mean SOC stock would increase by 0.040 kg m⁻², of which the largest proportion was contributed by oasis. The implementation of ecological restoration projects and salinity management in oasis has led to significant reduction in regional soil salinity and the areas of saline lands, which were estimated to additionally enhance SOC stock by 0.015 kg m⁻², accounting for about 10% of the total SOC stock increase over the study period. The proposed methods in this study can be applied for quantifying the direction and size of the salinity effect on SOC stock changes in other salt-affected regions, including costal salt marsh areas with high salinity and semi-humid agroecosystems with strong secondary salinization.

Our results also suggest that the effects of soil salinization should not be neglected in SOC modeling and future projection in arid inland river basins, especially for the oasis regions, and highlight the importance of soil salinization control in enhancing soil carbon sequestration. In the future, with the growing water resources being transferred from other external river basins in southwestern and southern China, more salts may be brought into the arid inland river basins of northwestern China after the water is consumed locally. Thus, effective measures should be further taken into practice to prevent soil salinization over the irrigated oasis of inland river basins. This could not only reduce the threat of soil salinization to both croplands and natural ecosystems but further enhance SOC stock as well. In addition, greater focus should also be placed on the interactions between salinity and other global change factors such as CO₂ enrichment, nitrogen decomposition, and climate extremes, and their co-effects on SOC dynamics in salt-affected inland regions.