Simulation and Spatio-Temporal Analysis of Soil Erosion in the Source Region of the Yellow River Using Machine Learning Method

Abstract

The source region of the Yellow River (SRYR), known as the “Chinese Water Tower”, is currently grappling with severe soil erosion, which jeopardizes the sustainability of its alpine grasslands. Large-scale soil erosion monitoring poses a significant challenge, complicating global efforts to study soil erosion and land cover changes. Moreover, conventional methods for assessing soil erosion do not adequately address the variety of erosion types present in the SRYR. Given these challenges, the objectives of this study were to develop a suitable assessment and prediction model for soil erosion tailored to the SRYR’s needs. By leveraging soil erosion data measured by ¹³⁷Cs from 521 locations and employing the random forest (RF) algorithm, a new soil erosion model was formulated. Key findings include that: (1) The RF soil erosion model significantly outperformed the revised universal soil loss equation (RUSLE) model and revised wind erosion equation (RWEQ) model, achieving an R² of 0.52 and an RMSE of 5.88. (2) The RF model indicated that from 2001 to 2020, the SRYR experienced an average annual soil erosion modulus (SEM) of 19.32 t·ha⁻¹·y⁻¹ with an annual total erosion in the SRYR of 225.18 × 10⁶ t·y⁻¹. Spatial analysis revealed that 78.64% of the region suffered low erosion, with erosion intensity declining from northwest to southeast. (3) The annual SEM in the SRYR demonstrated a downward trend from 2001 to 2020, with 83.43% of the study area showing improvement. Based on these findings, measures for soil erosion prevention and control in the SRYR were proposed. Future studies should refine the temporal analysis to better understand the influence of extreme climate events on soil erosion, while leveraging high-resolution data to enhance model accuracy. Insights into the drivers of soil erosion in the SRYR will support more effective policy development.

1. Introduction

Soil erosion significantly contributes to global land degradation, leading to the loss of soil nutrients and negatively impacting soil-related ecosystems, ultimately threatening regional ecological security [1]. The source region of the Yellow River (SRYR), located in the northeastern part of the Qinghai-Tibet Plateau [2], is often known as the “Chinese Water Tower”. It is a vital water source and an important ecological conservation area within the Yellow River Basin [3], serving as a crucial ecological barrier and grassland ecosystem in China [4]. However, due to the combined impacts of human activities and climate change, the SRYR suffers from widespread soil erosion, particularly erosion caused by water, wind, and freeze–thaw cycles [5], and rainfall erosion in this region has also intensified significantly since 1971 [6]. Additionally, the region experiences average wind speeds ranging from 0.6 to 4.2 m/s, which are significantly higher than in neighboring areas [7], thereby raising the potential for soil erosion. Studies have demonstrated that between 1950 and 2007, the average annual soil erosion modulus (SEM) in the SRYR stood at 18.68 t·ha⁻¹·y⁻¹, which represents a notable increase compared to pre-1950 levels [8]. Soil erosion has also exacerbated ecosystem degradation and led to a decline in water quality in the SRYR. Furthermore, it has adversely affected animal husbandry, which serves as the predominant source of income for the region’s residents, thus affecting their livelihoods. Consequently, there is an urgent requirement to accurately monitor intensity of soil erosion in the SRYR and establish a robust theoretical basis for the formulation of effective soil and water conservation strategies aimed at preventing and controlling soil erosion.

Soil erosion is a multifaceted process, driven by an array of interacting factors such as vegetation coverage, rainfall erosivity, soil texture, wind speed, freeze–thaw cycles, and human activities [5]. Current methods used to study soil erosion mainly include the runoff plot method, tracer methods, and modeling approaches [9]. The runoff plot technique, widely employed to ascertain the SEM of specific plots, while effective in yielding credible results on a small regional scale [10], faces challenges when estimating soil erosion over large scales, such as the SRYR, and does not provide long-term erosion data.

The nuclide tracer methods utilize radionuclides as tracers to delineate the research object. Among these, cesium-137 (¹³⁷Cs) is currently the most widely used and mature tracer technology [11,12]. As an artificial radionuclide, ¹³⁷Cs is deposited on the surface with water vapor and is immediately and strongly adsorbed by soil colloid particles. Given that Cs is a highly active alkali metal, positively charged cesium ions are easily adsorbed to negatively charged sites in soil organic matter, and essentially no further exchange reactions occur [13,14]. Therefore, the subsequent migration of ¹³⁷Cs is governed by the spatial flow and vertical movement of the soil, rather than chemical processes [15,16,17,18], making it a valuable tool for tracing soil erosion. This tracing technology has been widely employed to estimate soil erosion across various spatial scales, with applications by Cheikha et al. [19] in northeastern Tunisia, Porto et al. [20] in southern Italy’s catchment areas, Li et al. [8] in the Sanjiangyuan national park, and Lu et al. [21] in the Yushu Mountain area of China. Furthermore, Meusburger et al. [22] demonstrated that model-based methods often yield coarse-grained estimates of soil erosion, whereas fallout radionuclide (FRN) methods offer more reliable long-term assessments. Shao et al. [23] applied ¹³⁷Cs to evaluate the SEM of typical alpine meadows on the Qinghai-Tibet Plateau, revealing a negative correlation between the soil and water loss rates with vegetation coverage.

Given the limitations of runoff plot and tracer methods in terms of scalability and resource intensity [24], various soil erosion research models have been formulated. These include empirical models like the universal soil loss equation (USLE), the revised universal soil loss equation (RUSLE), and the revised wind erosion equation (RWEQ), as well as physical process models like the water erosion prediction project (WEPP) and the Limburg soil erosion model (LISEM) [25,26,27]. Among these, the RUSLE model and the RWEQ model are frequently utilized in large-scale studies because of their simplicity as well as empirical basis. However, as research deepens, several shortcomings have become apparent: (1) the accuracy of these models depends on site-specific parameters, which are typically calibrated to normal or moderate erosion. In reality, extreme erosion events are frequent, reducing the model’s accuracy [28]. (2) The linear assumptions inherent in these models fail to adequately capture complex physical processes under extreme climate conditions [28,29]. (3) The models consider only the inherent parameters of water and wind erosion dynamics, neglecting the interactions of multiple external forces, thereby overlooking compound erosion phenomena [30,31]. This is particularly pertinent in regions like the SRYR, where the variability in environmental factors is significant, including topography, climate, and vegetation. The SRYR encompasses nearly all types of soil erosion forces on land [32]. Therefore, verifying the applicability of classic models and identifying a model suitable for the SRYR is an urgent scientific challenge.

To overcome these challenges and enhance the precision of soil erosion assessments, researchers have increasingly turned to advanced machine learning techniques. Machine learning technology offers a promising approach by estimating SEM and examining intricate relationships between soil erosion and its contributing factors, proving highly effective for such assessments [33]. Among various algorithms, the random forest (RF) is particularly notable for constructing a diverse array of decision trees during the training process, which alternates the mode of classification classes or average prediction for regression tasks. This ensemble learning approach mitigates overfitting and bolsters generalization. By combining the outcomes of multiple decision trees, RF enhances model robustness and accuracy [34]. Given these characteristics, the RF model has demonstrated superior evaluation performance in assessing soil erosion across diverse geographic regions, including Central Asia [34,35,36,37], South Asia [38,39], northern Europe [40,41] hilly areas in southwestern and southern China [42,43], and the Tibetan Plateau [44].

In assessing soil erosion in the SRYR, the runoff plot method has proven insufficient for large-scale studies. The region encompasses nearly all forms of land-based soil erosion, and the traditional model’s applicability in this area necessitates further verification. Consequently, there is a critical necessity to establish a soil erosion assessment and prediction model tailored to the SRYR. This research compiled an environmental variable dataset comprising soil factors, meteorological factors, topographic factors, and remote sensing indices. The SEM measured by ¹³⁷Cs served as the verification basis. The aims of this study were to: (1) identify the most suitable soil erosion evaluation and prediction model by comparing the traditional empirical models with those based on the machine learning algorithms; (2) examine the spatial and temporal distribution of soil erosion in the study area from 2001 to 2020 utilizing the best-fit model and provide a solid foundation for developing soil and water conservation policies tailored in the SRYR.

2. Materials and Methods

2.1. Study Area

The SRYR (Figure 1) refers to the region extending from the source of the Yellow River to the Tangnaihai Hydrological Station, located from 32°30′−35°0′ N, 95°50′−103°30′ E, encompassing an area of roughly 12.2 × 10⁴ km² [45,46,47]. Positioned in the northeastern Qinghai-Tibet Plateau, it forms the headwaters of the Yellow River, Asia’s third-longest river. The SRYR experiences a plateau continental climate, characterized by decreasing temperatures and precipitation from east to west, with minimal annual temperature variations. The average annual temperature ranges from −1.7 °C to 3.4 °C [47], with temperatures typically above 0 °C from May to September. The average annual freezing period exceeds 160 days for many years [48]. The seasonal and interannual variations of precipitation are significant, with summer and autumn precipitation contributing over 75% of the annual total, averaging 491.8 mm [49]. The region’s elevation ranges from 2680 m to 6248 m, featuring a topography dominated by mountains, hills, and basins [50]. The land cover data utilized for this research combined global land cover data at a 30 m spatial resolution (accessed on 28 March 2024, https://www.webmap.cn/mapDataAction.do?method=globalLandCover) with a grassland type survey map of China from the 1980s in ArcGIS. The predominant land cover in the study area consists of grasslands, including alpine meadow (61.40%), alpine grassland (12.04%), and mountain meadow (9.29%), collectively covering 82.73% of the total area. Other land cover types include wetlands (6.56%), barren land (5.52%), and cultivated land (2.11%), with the remaining types covering areas less than 1.0% [51]. The primary land cover types in the SRYR are alpine meadow soil and alpine grassland soil, featuring layers and coarse textures. The soil composition is 14.5% clay, 39.90% silt, and 38.17% gravel [51].

2.2. Data Source and Processing

2.2.1. Measured Data Acquisition and Processing

The soil-measured data play a crucial role in constructing the suitable model in this research. The selection of sampling points for ¹³⁷Cs needed to represent general characteristics of the SRYR, while also considering practicalities and costs associated with sample collection. To tackle these challenges, the conditional Latin hypercube method (cLHS) [52], constrained by cost layers, was employed to arrange the sample sites. This was implemented by the cLHS tool in RStudio 1.3. A total of 521 soil samples containing ¹³⁷Cs were gathered across the SRYR between July and August in 2015, 2018, 2019, and 2020. The spatial configuration of these sampling sites is shown in Figure 1. A soil drill with a diameter of 5 cm was employed, targeting the 0–20 cm soil stratum, where over 98% of ¹³⁷Cs content resides [53]. At each site, 3 replicate samples were taken. The soil samples were then air-dried, ground, and sifted for further analysis. An N-type coaxial high-purity germanium probe, model GMX30P4, was employed for the measurements. Standard source comparison ensured accuracy, with the measurement duration set at 82,800 s. Typically, at a 95% confidence level, the analytical accuracy is approximately ±6%, with spectra interpreted using HPGe Gamma Spectrometer GCW 3523.

The methodology for calculating the SEM for these sampling sites was adapted from Wang et al. [53]. The calculation formula is as follows:

$$\mathrm{S}\mathrm{EM}={\displaystyle \frac{\mathrm{CPR}\times \mathrm{BD}\times \mathrm{DI}\times {10}^4}{\mathrm{T}}}$$

(1)

where SEM stands for soil erosion modulus (t·ha⁻¹·y⁻¹), CPR indicates change rate of ¹³⁷Cs during measurement period (%), BD refers to soil bulk density at the sampling sites (Mg·m⁻³), DI represents thickness of the soil layer (m), T denotes the time span from the peak of ¹³⁷Cs deposition (1963) to the sampling date.

CPR’s calculation formula is as follows:

$$\mathrm{CPR}={\displaystyle \frac{\left(\mathrm{CPI}-\mathrm{CRI}\right)\times 100}{\mathrm{CRI}}}$$

(2)

$$\mathrm{CPI}={\displaystyle \frac{\mathrm{Ci}\times \mathrm{W}}{\mathrm{s}}}$$

(3)

where CPI indicates ¹³⁷Cs concentration in the sample, CRI denotes ¹³⁷Cs reference inventory (Bq·m⁻²), Ci represents ¹³⁷Cs radioactivity (Bq·kg⁻¹), W refers to sample mass (kg), and s is cross-sectional area of the soil drill (m²).

CRI refers to atmospheric deposition flux of ¹³⁷Cs per unit area or the undisturbed ¹³⁷Cs area concentration value [54]. Accurately determining CRI is a prerequisite for using ¹³⁷Cs tracer technology to measure soil erosion. In this study, CRI was obtained using field exploration sampling and laboratory measurement methods. Typically, there is no erosion and accumulation at the top of the slopes [55]. Thus, the reference inventory site for this study was located at the flat part of the top of the slope, where the vegetation root layer is intact and compact. The sample reference inventory was collected at this location, resulting in 3 reference inventories: 2437.82 Bq·m⁻², 2225.42 Bq·m⁻², and 2199.61 Bq·m⁻². The mean values of these 3 was 2287.44 Bq·m⁻², which is consistent with previous studies [56,57].

2.2.2. Environmental Variable Data Acquisition and Processing

Soil erosion is a complex surface phenomenon affected by diverse environmental factors. When selecting variables that impact soil erosion in the SRYR, it is essential to take into account both the existing research basis and the unique characteristics of the SRYR, so as to thoroughly understand influencing mechanisms driving soil erosion in this region. Therefore, this study identified 4 categories of environmental factors, encompassing a total of 34 variables (detailed in Table S1). These factors include:

• Soil factors: 5 variables, including clay, sand, silt, soil organic carbon (SOC), and soil erodibility (K) factor;
• Meteorological factors: 8 variables, including precipitation, temperature, rainfall erosivity (R) factor, wind, illumination, freeze–thaw (FT), nighttime land surface temperature (LsTN), and daytime land surface temperature (LsTD);
• Topographic factors: longitude, latitude, elevation, slope, aspect, and length–slope (LS) factor;
• Remote sensing indices: 15 variables, including enhanced vegetation index (EVI), normalized difference vegetation index (NDVI), coverage, leaf area index (LAI), modified soil adjusted vegetation index (MSAVI), normalized difference vegetation green index (NDVGI), soil adjusted vegetation index (SAVI), difference vegetation index (DVI), normalized difference soil index (NDSI), normalized difference water index (NDWI), optimized soil adjusted vegetation index (OSAVI), ratio vegetation index (RVI), soil adjusted total vegetation index (SATVI), soil color index (SCI), transformed vegetation index (TVI), and coverage.

The data sources for each factor and some data processing methods are presented in

Table 1. Data sources and processing.

Variable Types	Data Sources	Data Processing
Soil factors	accessed on 1 April 2024 https://data.isric.org/	-
Meteorological factors	accessed on 1 April 2024 http://data.cma.cn https://data.tpdc.ac.cn/zh-hans/	Meteorological station point data were interpolated into raster data format with a 1 km spatial resolution using ANUSPLIN tool.
Topographic factors	accessed on 15 April 2024 http://srtm.csi.cgiar.org/	Calculated using the DEM and adjusted to a 1 km spatial resolution.
Remote sensing indices	accessed on 28 April 2024 https://code.earthengine.google.com/	-

.

2.3. Methods and Modeling

2.3.1. RUSLE Model

The revised universal soil loss equation (RUSLE) is a frequently applied method for estimating water erosion [58]. According to Zhao et al. [59], the calculation formula is given as:

$$\mathrm{A}=\mathrm{R}\times \mathrm{K}\times \mathrm{LS}\times \mathrm{C}\times \mathrm{P}$$

(4)

where A denotes hydraulic erosion modulus (t·ha⁻¹·y⁻¹), R refers to rainfall erosivity factor (Mj·mm·ha⁻¹·h⁻¹·y⁻¹), K represents soil erodibility factor (t·ha·h·Mj⁻¹·mm⁻¹), LS stands for the length and slope factor, C refers to the vegetation cover management factor, P indicates soil and water conservation practices.

The calculation formulas for the R, K, and C factors are as follows:

$$\mathrm{R}=\sum _{\mathrm{j}=1}^{12}1.735\times 10^{\left[\left(1.5\times \lg {\displaystyle \frac{{\mathrm{p}}_{\mathrm{j}}^2}{\mathrm{p}}}\right)-0.8188\right]}$$

(5)

$$\begin{gathered} \mathrm{K}=\left\{0.2+0.3\exp \left[-0.0256\mathrm{Sand}\left(1-{\displaystyle \frac{\mathrm{Silt}}{100}}\right)\right]\right\}\times {\left({\displaystyle \frac{\mathrm{Silt}}{\mathrm{Clay}+\mathrm{Silt}}}\right)}^{0.3}\times \\ \begin{array}{c}\left[1-{\displaystyle \frac{0.25\mathrm{SOC}}{\mathrm{SOC}+\exp \left(3.72-2.95\mathrm{SOC}\right)}}\right]\times \left[1-{\displaystyle \frac{0.7{\mathrm{SN}}_1}{\mathrm{S}{\mathrm{N}}_1+\exp \left(-5.51+22.9\mathrm{S}{\mathrm{N}}_1\right)}}\right]\end{array} \end{gathered}$$

(6)

$$\mathrm{S}{\mathrm{N}}_1=1-{\displaystyle \frac{\mathrm{Sand}}{100}}$$

(7)

$$\mathrm{C}=\left\{\begin{array}{ll}1& \mathrm{VC}=0\\ 0.6508-0.3436\lg \left(\mathrm{VC}\right)& 0\text{<}\mathrm{VC}\le 78.3\%\\ 0& \mathrm{VC}\text{>}78.3\%\end{array}\right.$$

(8)

$$\mathrm{VC}={\displaystyle \frac{\mathrm{NDVI}-\mathrm{NDV}{\mathrm{I}}_{\min }}{\mathrm{NDV}{\mathrm{I}}_{\max }-\mathrm{NDV}{\mathrm{I}}_{\min }}}$$

(9)

where p_j represents monthly rainfall (mm), and p indicates annual rainfall (mm). Sand, Silt, Clay, and SOC represent the respective content percentages in the soil. VC denotes vegetation cover. NDVI was obtained from Google Earth Engine (GEE).

The LS factor was derived from the DEM. The P factor was assigned according to land cover type. Considering that grassland accounts for 82.73% of the study area and human impacts are limited, the P factor was set to 1 [60].

2.3.2. RWEQ Model

The revised wind erosion equation (RWEQ) is a frequently utilized model for estimating wind erosion [61]. According to Teng et al. [62], its calculation formulas are provided as follows:

$${\mathrm{S}}_{\mathrm{L}}={\displaystyle \frac{2\mathrm{z}}{{\mathrm{S}}^2}}{\mathrm{Q}}_{\max }{\mathrm{e}}^{-{\left({\displaystyle \frac{\mathrm{z}}{\mathrm{S}}}\right)}^2}$$

(10)

where S_L denotes wind erosion modulus (t·ha⁻¹·y⁻¹), z indicates maximum downwind erosion distance (m), S represents potential regional erosion coefficient, Q_max signifies maximum potential wind erosion transfer (kg·m⁻¹).

S and Q_max’s calculation formulas are as follows:

$$\mathrm{S}=150.71(\mathrm{WF}\times \mathrm{EF}\times \mathrm{SCF}\times {\mathrm{K}}^{\prime }\times {\mathrm{C}}_1{)}^{-0.3711}$$

(11)

$${\mathrm{Q}}_{\max }=109.8(\mathrm{WF}\times \mathrm{EF}\times \mathrm{SCF}\times {\mathrm{K}}^{\prime }\times {\mathrm{C}}_1)$$

(12)

where WF represents meteorological factor (kg·m⁻¹), EF indicates soil erodibility factor (t·ha·h·Mj⁻¹·mm⁻¹), SCF denotes soil crust factor, ${\mathrm{K}}^{\prime }$ is surface roughness factor, and C₁ is the comprehensive vegetation factor.

WF’s calculation formulas are as follows:

$$\mathrm{WF}=\mathrm{Wf}\times {\displaystyle \frac{\mathsf{\rho }}{\mathrm{g}}}\times \mathrm{SW}\times \mathrm{SD}$$

(13)

$$\mathrm{Wf}={\mathrm{u}}_2({\mathrm{u}}_2-{\mathrm{u}}_1{)}^2\times {\mathrm{N}}_{\mathrm{d}}$$

(14)

where Wf denotes wind factor (m³·s⁻³), ρ indicates air density (kg·m⁻³), g refers to acceleration due to gravity (m·s⁻²), SW represents soil moisture factor, SD represents snow cover factor, u₂ denotes interpolated wind speed (m·s⁻¹), u₁ denotes sand-raising wind speed (m·s⁻¹), and Nd represents the number of days with wind speed exceeding 5 m·s⁻¹.

EF and SCF’s calculation formulas are as follows:

$$\mathrm{EF}={\displaystyle \frac{29.09+0.31{\mathrm{s}}_{\mathrm{a}}+0.17{\mathrm{s}}_{\mathrm{i}}+0.33(\frac{{\mathrm{s}}_{\mathrm{i}}}{{\mathrm{c}}_1})-2.59\mathrm{OM}-0.95{\mathrm{CACO}}_3}{100}}$$

(15)

$$\mathrm{SCF}={\displaystyle \frac{1}{1+0.0066{\mathrm{c}}_1^2+0.021\mathrm{O}{\mathrm{M}}^2}}$$

(16)

where s_a, s_i, c₁, OM, and CACO₃ represent the respective contents of sand, silt, clay, organic matter, and calcium carbonate in the soil (%).

${\mathrm{K}}^{\prime }$ and C₁’s calculation formulas are as follows:

$${\mathrm{K}}^{\prime }=\cos \alpha$$

(17)

$${\mathrm{C}}_1={\mathrm{e}}^{-0.0483\mathrm{Coverage}}$$

(18)

where α denotes the terrain slope, and Coverage indicates the vegetation cover.

2.3.3. Machine Learning Model

There is a high probability of multicollinearity between the environmental variables set previously. Therefore, it is necessary to screen variables before building a model to solve the multicollinearity problem, avoid model complexity caused by information redundancy, and improve model accuracy and computational efficiency [63]. The LASSO algorithm, a linear regression technique used for feature selection, constructs a penalty function to eliminate variables with zero coefficients. This process minimizes the sum of squares of the model residuals, ultimately achieving effective feature selection [64]. The LASSO algorithm demonstrates superior performance when processing high-dimensional remote sensing data [65]. In this study, LASSO was employed to address the issue of multicollinearity among variables.

In recent years, various machine learning algorithms, such as random forest (RF), support vector machine (SVM), K-nearest neighbor (KNN), boosted regression tree (BRT), and extreme gradient boosting (XGBoost), have gained significant traction in research fields like soil erosion and gully erosion [66]. RF, an advanced variant of the decision tree model, leverages advanced non-parametric techniques to surpass the accuracy of other machine learning algorithms [34]. It is recognized as one of the leading machine learning approaches [67,68] and has been extensively utilized for evaluating and predicting soil organic matter, erosion, bedrock depth, as well as soil physical and chemical properties [69]. Two training parameters needed to be set for modeling using the random forest algorithm: the number of regression trees was set to the default value of 500, and the number of randomly selected predictors at each node was set to the square root of the total number of predictors. Then, the relationship between the measured SEM of ¹³⁷Cs and the erosion influencing factors was modeled to estimate the SEM across the SRYR. The Gini coefficient is used to evaluate the importance of each impacting factor in the RF model. Specifically, each factor’s importance is qualified by calculating impurity reduction when splitting the node [70]. This method aids in identifying the most influential variables affecting soil erosion in the SRYR.

Model performance was assessed through a ten-fold cross-validation approach [70]. Here, 90% of dataset was employed for training and 10% for validation. The R² and RMSE were computed for each fold. The RF model’s performance was evaluated by averaging the R² and RMSE across both training and validation sets over ten iterations.

2.4. Trend Analysis

The Theil–Sen (SEN) median trend analysis and Mann–Kendall test are non-parametric statistical test methods commonly applied in long-term time series data trends [71,72]. Compared with traditional linear trend analysis methods, these approaches offer higher computational efficiency and robustness. Therefore, these methods were employed to determine the soil erosion change trend and significance in the SRYR. When ${\mathrm{S}\mathrm{e}\mathrm{n}}_{\mathrm{s}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}$ > 0, it indicates an increasing trend of SEM, reflecting a worsening condition of soil erosion. Conversely, a ${\mathrm{S}\mathrm{e}\mathrm{n}}_{\mathrm{s}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}}$ < 0 implies an improvement in soil erosion status. The trend’s significance was tested at the α = 0.05 confidence level. A z > |1.96| signifies a significant change trend (α < 0.05), while z $\le$ |1.96| suggests no significant change (α < 0.05). The evaluation standards are shown in

Table 2. Standards for assessing change trends in soil erosion in the SRYR.

${\mathbf{S}\mathbf{e}\mathbf{n}}_{\mathbf{s}\mathbf{l}\mathbf{o}\mathbf{p}\mathbf{e}}$	Z Value	Trends in Soil Erosion
≥0.001	≥1.96	Significantly worsened
≥0.001	−1.96–1.96	Slightly worsened
−0.001–0.001	−1.96–1.96	Stable
≤−0.001	−1.96–1.96	Slightly improved
≤−0.001	≤−1.96	Significantly improved

.

3. Results

3.1. RUSLE Model Results

The scatter fitting diagram (Figure 2d) reveals that the RUSLE model results did not fit well with the SEM calculated by ¹³⁷Cs (R² = 0.0082). The SEM in the SRYR could be classified based on the soil and water loss classification standard (Table S2) issued by the Ministry of Water Resources of the People’s Republic of China in 2007 [73]. The annual average SEM in the SRYR calculated by the RUSLE model was 14.81 t·ha⁻¹·y⁻¹, contributing to an annual total erosion of 169.04 × 10⁶ t·y⁻¹, suggesting that the SRYR experienced low erosion. Specifically, in the northwest and southeast of the SRYR, soil erosion volume was minimal, whereas it was higher in the central areas. The SEM was notably high in Maqin, Dari, Gande, Maqu, Jiuzhi, and Xinghai. The annual average SEM in Jiuzhi, Maqin, and Dari reached 34.51 t·ha⁻¹·y⁻¹, 24.14 t·ha⁻¹·y⁻¹, and 22.83 t·ha⁻¹·y⁻¹, categorizing these areas as moderate-, low-, and low-erosion zones. Overall, the estimation results of the RUSLE model indicated that predominant erosion types in the SRYR were classified as slight and low erosion, constituting 84.35% and 15.57% of the SRYR.

3.2. REWQ Model Results

The scatter fitting diagram (Figure 2e) reveals that the RWEQ model results did not align closely with the SEM calculated by ¹³⁷Cs well (R² = 0.0301). The annual average SEM for the SRYR calculated by the RWEQ model was 24.93 t·ha⁻¹·y⁻¹, contributing to a total soil loss of 282.83 × 10⁶ t·y⁻¹, indicating that the SRYR experienced low erosion. The evaluation results of the RWEQ model were generally higher than those of the RUSLE model and the RF model, with a more dispersed spatial pattern. The RWEQ model identified low and moderate erosion as the predominant types in the SRYR, constituting 44.56% and 51.64% of the SRYR, respectively. Low erosion was mainly concentrated in Maqin, moderate erosion was spread throughout the SRYR, and high erosion (0.7%) was concentrated at the junction of Xinghai and Tongde.

3.3. RF Model Results

3.3.1. Variable Selection

The LASSO method was used to select eleven environmental variables, including two soil factors (SOC and clay), three meteorological factors (rainfall, wind, and illumination), two topographic factors (latitude and elevation), and four remote sensing indices (LAI, NDVI, NDSI, and OSAVI). Figure S1 depicts the distribution of the average values of 11 variables over the period from 2001 to 2020. Among these eleven variables, latitude, elevation and clay do not change over time, while the remaining eight variables have values recorded from 2001 to 2020. In total, 163 datasets were applied to the RF algorithm across different years to build a soil erosion evaluation and prediction model.

3.3.2. Model Results and Accuracy Evaluation

As seen from the scatter fitting diagram (Figure 2f), the R²_train value for the training set was 0.9, close to 1, and the RMSE for the training set was 2.95, indicating that the RF model’s regression performance was robust and the model was deemed reliable. The annual average SEM in the SRYR calculated by the RF model was 19.32 t·ha⁻¹·y⁻¹, contributing to a total erosion volume of 5.18 × 10⁶ t·y⁻¹, indicating low erosion in the SRYR. The RF model’s classification results for soil erosion in the SRYR were consistent with those from the RUSLE model and the RWEQ model. However, SEM calculated by the RF model fell between those of the RUSLE and the RWEQ model. The SEM in the southeast was low, while it was high in the northwest, indicating a decrease in erosion intensity from northwest to southeast. The SRYR was mainly characterized by low erosion as well as moderate erosion, covering 78.64% and 21.36% of the study area, respectively.

3.4. Spatial Distribution of Soil Erosion in the SRYR

The RF model outperformed the other two models, allowing for a detailed analysis of the spatial and temporal patterns of soil erosion in the SRYR using its estimation results. To effectively utilize the soil erosion simulation outcomes for management purposes, the results were disaggregated by county (

Table 3. SEM, total erosion volume, and proportion for each county in the SRYR between 2001 and 2020.

County	Max (t·ha−1·y−1)	Min (t·ha−1·y−1)	Mean (t·ha−1·y−1)	Total (106 t·y−1)	Proportion (%)
Maduo	33.52	14.62	25.63	52.18	23.17
Maqin	33.03	14.40	19.06	25.36	11.26
Qumalai	33.60	15.18	28.89	22.90	10.17
Dari	32.30	14.09	17.69	20.46	9.09
Maqu	30.73	13.57	15.53	14.75	6.55
Gande	32.90	14.35	16.45	12.02	5.34
Chindu	33.04	15.06	23.76	10.91	4.84
Xinghai	32.76	14.74	19.97	10.59	4.70
Ruoregai	23.05	12.73	14.96	10.12	4.49
Hongyuan	28.98	12.08	13.58	9.14	4.06
Jiuzhi	32.05	13.44	15.19	9.11	4.05
Tongde	32.56	14.59	17.96	8.13	3.61
Henan	25.04	14.51	15.75	7.75	3.44
Zeku	32.17	14.79	16.23	6.87	3.05
Aba	20.47	12.79	14.04	4.90	2.18

). Among the 15 counties in the SRYR, Qumalai, Maduo, and Chindu reached the highest annual average SEM, at 28.89 t·ha⁻¹·y⁻¹, 25.63 t·ha⁻¹·y⁻¹, and 23.76 t·ha⁻¹·y⁻¹. Both Qumalai and Maduo counties were classified as moderate-erosion areas. Moreover, Maduo, Maqin, and Qumalai counties had the highest total soil erosion, at 52.18 × 10⁶ t·y⁻¹, 25.36 × 10⁶ t·y⁻¹, and 22.9 × 10⁶ t·y⁻¹. Although Chindu had the highest SEM, only part of this county is located in the SRYR, resulting in a relatively low total erosion amount. Conversely, the counties with the lowest mean SEM were Hongyuan, Aba, and Ruoergai, at 13.58 × 10⁶ t·y⁻¹, 14.04 × 10⁶ t·y⁻¹, and 14.04 × 10⁶ t·y⁻¹. The NDVI in these areas was relatively high (Figure S1i), indicating good vegetation coverage. This robust vegetation slowed the scouring effect of water on the soil, while plant roots stabilized the soil, thereby reducing erosion caused by both rain and wind [74].

To assess how different cover types resist soil erosion, six main land cover types were selected for analysis (

Table 4. SEM, total erosion volume and proportion for main land cover types in the SRYR between 2001 and 2020.

Land Cover Type	Max (t·ha−1·y−1)	Min (t·ha−1·y−1)	Mean (t·ha−1·y−1)	Total (106 t·y−1)	Proportion (%)
Alpine steppe	33.52	12.08	24.74	34.68	15.64
Alpine meadow	33.50	12.23	18.49	135.39	61.08
Mountain meadow	25.95	12.14	14.31	15.60	7.04
Cropland	32.53	13.43	17.10	4.13	1.86
Wetland	33.35	12.34	16.98	12.70	5.73
Barren	33.60	14.25	28.15	17.83	8.04

): alpine grassland, alpine meadow, mountain meadow, cropland, wetland, and barren. Among these, barren land exhibited the highest annual average SEM at 28.15 t·ha⁻¹·y⁻¹, mountain meadow recorded the lowest at 14.31 t·ha⁻¹·y⁻¹. Between 2001 and 2020, the alpine grassland demonstrated both the peak SEM value at 33.52 t·ha⁻¹·y⁻¹ and the lowest value at 12.08 t·ha⁻¹·y⁻¹. Furthermore, the alpine meadow suffered the highest total soil erosion, reaching 135.39 × 10⁶ t·y⁻¹.

3.5. Temporal Distribution of Soil Erosion in the SRYR

The SEM in the SRYR from 2001 to 2020 ranged between 18.04 t·ha⁻¹·y⁻¹ and 20.12 t·ha⁻¹·y⁻¹, with the highest and lowest values recorded in 2003 and 2018, respectively (Figure 3). During this period, the SEM followed a declining trend. The area of moderate soil erosion decreased by 9833 km², while the low-erosion area increased by the same amount. The SEM notably decreased from 2003 to 2005, likely due to the intensive implementation of Chinese government policies, including the “Grain to Green Program”, and the soil and water conservation policies initiated in the SRYR since 2003 [75]. From 2006 to 2015, SEM exhibited an up-and-down trend but dropped sharply from 2015 to 2018. This sharp decline may be linked to the second round of grassland ecological compensation policy implemented by the Chinese government in 2016, which increased subsidies for herders to ban grazing, thus alleviating pressure on grasslands and promoting the restoration of grassland vegetation [76]. The trends and fluctuations of SEM in 15 counties from 2001 to 2020 were consistent with those in the overall SRYR (Figure 4). All counties showed varying degrees of decline around 2016. The downward trend was most pronounced in Hongyuan, Aba, and Chindu. Conversely, the fluctuation trend was smaller in Hongyuan, Ruoergai, and Tongde, while larger fluctuations were observed in Jiuzhi, Qumalai, and Henan. The changing trends and fluctuations of SEM across the six main land cover types from 2001 to 2020 were also consistent with those of the SRYR as a whole (Figure 5). Among these, the downward trend of SEM was most evident in alpine grassland, cropland, and mountain meadow. The fluctuation trend was smallest in cropland and largest in barren land.

The dynamic trend of soil erosion in the SRYR between 2001 and 2020 (Figure 6) was determined using the Sen trend analysis in conjunction with the Mann–Kendall test. Over the last 20 years, the soil erosion conditions have generally improved in the SRYR. The improved area accounted for 83.43% of the entire study area, with 32.61% of the area showing significant improvement, mainly in the southeast. The slightly improved area accounted for 50.82% and was spread throughout the entire study area. Only 2.07% of the SRYR remained stable, primarily distributed in the northwestern region. Conversely, the area experiencing worsening conditions accounted for 14.5%, with 13.5% undergoing slight worsening, particularly in the northwest of Qumalai, the south of Xinghai, and the east of Maduo and Gande. The significantly worsened areas made up 0.99%, mostly concentrated in the central part of Xinghai.

From the perspective of county administrative divisions (Figure S2a), Hongyuan (93.65%), Aba (86.26%), and Zeku (89.48%) had a significantly higher proportion of improved areas (including both slightly and significantly improved) compared to other counties and the overall average level of the SRYR, which stood at 32.61%. Similarly, counties like Xinghai (8.70%), Qumalai (1.86%), Maqin (1.47%), and Chindu (1.39%) exhibited a higher proportion of worsened areas (including both slightly and significantly worsened) than other counties and the overall SRYR average of 0.99%. Counties where improved soil erosion conditions accounted for over 90% of their total area include Aba (99.69%), Henan (99.59%), Ruoergai (99.49%), Hongyuan (98.9%), Maqu (98.29%), Jiuzhi (98.21%), and Zeku (94.1%). Meanwhile, counties with improved areas covering 80% to 90% of their land include Dari (85.3%), Maduo (81.44%), and Gande (80.36%). Notably, the proportion of worsened soil erosion exceeded 30% in Xinghai (53.00%) and Qumalai (33.04%), marking these regions as critical zones for erosion control.

From a land cover perspective (Figure S2b), mountain meadow (71.63%) and wetland (64.92%) had significantly higher proportions of improved areas compared to other land cover types and the SRYR average. In contrast, barren land (1.72%), alpine grassland (1.10%), and alpine meadow (1.04%) showed a higher proportion of significantly worsened areas compared to other land cover types and the SRYR average. The order of soil erosion improvement, from highest to lowest, is as follows: mountain meadow (97.93%), wetland (95.21%), cropland (94.74%), alpine meadow (82.55%), alpine grassland (78.85%), and barren land (63.83%).

4. Discussion

4.1. Model Comparisons and Effectiveness

From the RUSLE model evaluation and the fitting scatter plot of the SEM calculated by ¹³⁷Cs (Figure 2d), it is evident that the model performed poorly for soil erosion in the SRYR (R² = 0.0082). The possible reasons are as follows: firstly, the RUSLE model operates as a linear model. According to the calculation formula of the C factor (Figure S3a), when vegetation coverage (VC) is greater than 78.3%, the C factor result is 0. The VC in the southeastern SRYR (Figure S1i) was relatively high. Therefore, when the RUSLE model was applied to estimate the SEM in the SRYR, the fitting results for Hongyuan, Aba, and Ruoergai in the southeast were close to 0 t·ha⁻¹·y⁻¹. Notwithstanding, soil loss still occurs under heavy rainfall or extreme climate conditions, despite high vegetation coverage. Secondly, the LS factor (Figure S3b) was minimal and close to 0 in the northwest, which also made the SEM in Maduo and Qumalai close to 0 t·ha⁻¹·y⁻¹. The LS factor calculation assumes uniform slope, neglecting slope shape and terrain complexity, contributing to model error [77]. Additionally, R factor (Figure S3c) in this study was found to be negatively correlated with soil erosion, which contradicts the theoretical basis of the RUSLE model. The discrepancy might be due to utilizing monthly rainfall data rather than cumulative heavy rainfall data for calculating the R factor.

From the fitting scatter plot of the RWEQ model evaluation results and the SEM estimated by ¹³⁷Cs (Figure 2e), it is evident that the performance (R² = 0.0301) was better than that of the RUSLE model, but still poor. In this study, annual rather than monthly meteorological data were used in the RWEQ model for evaluation, failing to capture the monthly variations in meteorological factors [78], which is the primary reason for the model’s deviation. Additionally, a study on wind erosion in the Tarim Basin of China and the Columbia Plateau of the United States indicated that even after calibrating the empirical parameters of the RWEQ model, its performance remained poor. This was likely due to the model underestimating the maximum potential wind erosion [79].

The SEM calculated by ¹³⁷Cs was used to establish and verify a custom soil erosion evaluation and prediction model, and the findings revealed that, in the SRYR, the soil erosion types primarily consisted of low erosion and moderate erosion. Between 2001 and 2020, the proportion of moderate erosion decreased, while the corresponding share of low erosion increased. These findings align with the conclusions of Li et al. [80], Li et al. [81], and Yin et al. [82]. Soil erosion peaked in the northwest and gradually diminished toward the southeast, consistent with the findings of Xiao et al. [83]. Moreover, sediment serves as a crucial indicator for assessing land degradation and erosion within a watershed. It is driven by factors such as rainfall, runoff, and other processes that erode soil particles from the surface [84]. The findings of this study are validated against monitoring data from the Tangnaihai Hydrological Station, situated downstream of the SRYR. In 2020, the station recorded an annual sediment delivery of 18.8 × 10⁶ t·y⁻¹ [85], while total erosion in the SRYR estimated by the RF model is 214.4 × 10⁶ t·y⁻¹, resulting in a sediment delivery ratio (SDR) of 0.088. The SDR aligns closely with values calculated in previous studies for the Tangnaihai Hydrological Station [86]. Thus, the custom soil erosion model established is considered to be more reliable.

4.2. Influence Factors

The RF algorithm’s feature importance results highlighted NDVI as the most significant factor influencing soil erosion in the SRYR (Figure S4), aligning with the findings by Wang et al. [87] and Yin et al. [82]. This research reported a negative correlation coefficient of −0.46 between NDVI and the SEM by ¹³⁷Cs. The higher the NDVI, the lower the SEM. NDVI, OSAVI, and LAI can all reflect the coverage and reproduction of vegetation. The OSAVI and LAI were also negatively correlated with the SEM, both showing a correlation coefficient of −0.42. In general, the surface layer of vegetation can effectively intercept and dissipate raindrops, and the root system can prevent the generation of surface runoff, thereby improving the resistance of soil to water erosion [88]. Meanwhile, vegetation coverage can reduce wind erosion by reducing wind speed and fixing sand and dust [89]. Precipitation ranked second in the importance of features in the RF algorithm (Figure S4). In this study, it was negatively correlated with the SEM calculated by ¹³⁷Cs (−0.37). This observation aligns with results from Zhang et al. [90] in central Oklahoma and Jobin Thomas et al. [91] in the Pamba River Basin (PRB), India. While many studies have pointed out that precipitation is positively correlated with SEM [92,93], the relationship between rainfall and soil erosion exhibited complexity. Appropriate rainfall helps to form the surface structural sealing effect of the soil and reduces the SEM [94]. However, it is worth noting that the intensity and duration of rainfall influence how precipitation affects soil erosion. The rainfall intensity and duration in the SRYR are smaller than those in southern China, where there are continuous and concentrated rainstorms. Low-intensity precipitation promotes soil moisture infiltration, reduces surface runoff [95], and provides favorable natural conditions for vegetation growth, slowing down soil erosion. Therefore, future research should incorporate analysis on a monthly scale, focusing on extreme climatic events like storms, to further understand rainfall’s impact on soil erosion [96], and use the heavy rainfall accumulation to characterize rainfall erosivity. Although altitude was not identified as the most significant factor for soil erosion in this study (Figure S4), it is crucial to recognize that its importance can vary in other studies. The impact of soil erosion factors often depends on regional characteristics and the scale of research [97]. The 30 m resolution DEM data employed here may have limited the depiction of large-scale terrain changes, affecting altitude’s role in explaining erosion. Future research should investigate the influence of terrain factors at varying spatial resolutions for a more comprehensive understanding.

A positive correlation between altitude and SEM was also found, consistent with findings by Nearing et al. [98]. This may be due to vegetation coverage being the most crucial factor affecting soil erosion in the SRYR. High-altitude areas have limited vegetation growth due to climatic conditions, making soil more prone to direct erosion from runoff and raindrops [99]. Li et al. found that altitude was a key factor influencing gully erosion in the Lhasa River Basin, where higher altitude was linked to a lower risk of gully erosion, attributed to larger catchment areas as well as increased kinetic energy of runoff at lower elevations [100]. Finally, soil components like SOC and clay contribute significantly by promoting soil stability and enhancing water retention, thus mitigating the effects of water flow [101]. NDSI, indicative of soil moisture, organic content, and structure, is inversely related to SEM. Further research should delve into how these soil properties influence erosion through empirical fieldwork and lab-based studies.

4.3. Implications for Soil Erosion Prevention and Control in the SRYR

The spatial and temporal patterns of soil erosion in the SRYR provide critical insights for developing soil erosion prevention and control strategies. Firstly, the magnitude of soil erosion and its trends should serve as the basis for classification, enabling prioritized management across counties with zoning and hierarchical strategies. Maduo, Chindu, and Qumalai counties require more intensive soil erosion control, with average SEM values of 28.89 t·ha⁻¹·y⁻¹, 25.63 t·ha⁻¹·y⁻¹, and 23.76 t·ha⁻¹·y⁻¹. Additionally, Qumalai and Chindu counties had significantly worsened soil erosion areas larger than the SRYR average (0.99%), at 1.86% and 1.39, respectively. They showed lower NDVI, OSAVI, and LAI values, higher NDSI, lower vegetation coverage, poorer soil conditions, and less precipitation compared to other counties (Figure S1). Consequently, revegetation measures are essential for erosion mitigation. Specifically, strict adherence to grazing bans during the greening period from early May to late June each year, the most fragile time for the SRYR ecosystem [102], can enhance pasture growth and increase vegetation coverage and biomass. In economically better-off areas, initiatives such as water diversion projects and artificial rainfall can aid in irrigating grasslands and restoring vegetation. Conversely, Hongyuan, Zeku, and Aba counties had lower soil erosion control needs, with the lowest annual average SEM values at 12.08 t·ha⁻¹·y⁻¹, 12.73 t·ha⁻¹·y⁻¹, and 12.79 t·ha⁻¹·y⁻¹ (Table 3). Their proportion of areas with significantly improved soil erosion was much higher than that of other counties in the SRYR, at 93.65%, 89.48%, and 86.26%, respectively (Table S4). For these regions, efforts should focus on enhancing grassland productivity by selecting grass species with strong climate adaptability and high yields, adhering to grass–livestock balance standards, and reducing grazing pressure on natural grasslands. The other nine counties had moderate soil erosion control demands. In these regions, the optimal grazing rate can be determined based on the pasture yields in low-yield years. Additionally, in regions with higher altitudes and severe rain erosion, biological measures and engineering measures can be implemented. Combined with small watershed management projects, these measures can mitigate losses caused by wind and water erosion.

Secondly, counties can implement soil erosion control measures tailored to their predominant land cover types within their regions. The annual average SEM on barren land was the highest, at 28.15 t·ha⁻¹·y⁻¹, with significantly worsened soil erosion on barren land accounting for 1.72% of the total study area (Table S4), making it the most impacted land cover type. There is a significant amount of barren land in northern Qumalai and western Maqin (Figure 1). To mitigate soil erosion, these areas should be strategically reclaimed, particularly in the relatively flat river valleys [103]. Finally, this research highlighted the effectiveness of the Chinese government’s soil conservation and ecological restoration initiatives in reducing erosion in the study area. SEM in the 15 counties declined after 2003 but rebounded slightly after 2005 (Figure 4), indicating the successful impact of “Grain to Green Program” since its intensive launch in 2003. Moreover, SEM decreased sharply around 2016 (Figure 4), reaching its lowest value in the entire study period around 2018. This decline may be attributed to the grassland ecological compensation policy implemented by the Chinese government in 2016. Continuing these efforts with consistent policy implementation is vital for sustainable long-term benefits.

5. Conclusions

This research led to the following conclusions:

(1) The RF-based custom soil erosion model outperforms the classical RUSLE model and the RWEQ model in evaluating the soil erosion in the SRYR.

(2) The SRYR experiences distinct spatial and temporal patterns of soil erosion, with erosion intensity diminishing from the northwest to the southwest. Over the period from 2001 to 2020, there has been a general decrease in soil erosion levels.

(3) NDVI emerged as the most critical factor in the RF algorithm, greatly influencing the SEM in the SRYR.

(4) Implementing zoning and hierarchical management strategies for soil erosion prevention and control is essential across the SRYR’s counties. Furthermore, barren land can be reclaimed effectively to enhance soil and water conservation.

Although numerical differences exist between the RF soil erosion model’s results and classic models, the temporal trends and spatial patterns of soil erosion in the SRYR have been validated in related studies. Moving forward, improving accuracy will require the use of higher-precision environmental data and refined model parameters. Additionally, it will be essential to examine the impact of extreme climate events on soil erosion and to expand the research scale from annual to quarterly and monthly intervals. Investigating the driving factors of soil erosion in the SRYR will provide a more robust theoretical foundation for policy formulation.