Driving factors of runoff and sediment yield effects of various slope ecological restoration measures in the middle reaches of the Yellow River

Introduction

Restoring the Yellow River Basin and managing its sustainable development is one of China’s major development projects (Zuo et al., 2023). The frequent occurrence of soil erosion disasters in the middle reaches of the Yellow River profoundly impacts the river’s sediment discharge (Mbezele et al., 2022). Liu et al. (2024b) used analytical tools such as the Mann–Kendall (M–K) trend test and double mass curve analysis to examine erosion patterns in these rivers from 1950 to 2022, as well as the primary drivers of these changes. Wang & Sun (2021) pointed out that the middle reaches contribute over 90% of the sediment deposited in the lower reaches of the Yellow River and 30% of its annual river discharge. Coarse sediment causes the uncoordinated relationship between runoff and sediment in the Yellow River, which makes the local ecological environment fragile. Therefore, it is important to study the relationship between runoff and sediment (Wang et al., 2022). The Miaochuan Basin of the middle reaches of the Yellow River, located in the Mu Us Sandy Land, is an arsenopyrite sandstone area that suffers from severe soil erosion due to the soil type (Yuan et al., 2025). China is highly concerned about the ecological and environmental issues in the middle reaches of the Yellow River. Since the 1980s, it has launched a series of projects for ecological conservation, such as the Three-North Shelterbelt Forest Program, the Project for Returning Farmland to Forests and Grassland and the Project for Side Slope Protection. These initiatives have significantly contributed to the prevention and control of soil erosion in the middle reaches of the Yellow River (Liu et al., 2025). However, the mechanism of runoff and sediment yield under various vegetation ecological restoration measures in this area remains unclear. Yin et al. (2024) showed that runoff depth (RD) and sediment yield (SY) are mainly influenced by precipitation, vegetation, landform and ecological restoration measures. The influence of rainfall on the water erosion process has great spatial and temporal heterogeneity, which is related to the characteristic factors such as rainfall, rainfall intensity, and rainfall duration (T) (Wu et al., 2017). Wischmeier & Smith (1958) pointed out that the product of rainfall kinetic energy (E) and maximum intensity of precipitation in thirty minutes (I₃₀) serves as a measure of precipitation erosivity. This erosivity index is used as an input parameter in the development of the Universal Soil Loss Equation (USLE), a model for predicting soil erosion caused by rainfall. According to Bochet, Poesen & Rubio (2018), there are complex relations between precipitation and runoff and both runoff depth and sediment yield are subject to precipitation intensity. Hamed et al. (2002) and Martinez-Murillo et al. (2013) found that runoff depth and sediment yield change with precipitation intensity, when other conditions remain relatively constant. Vegetation and topography have strong regional characteristics. Vegetation has the effects of increasing surface coverage, reducing flood peaks, and fixing soil (Li et al., 2016). Slope steepness (S) and slope length are two factors related to the implications of landform on soil erosion (Liu, Zhang & Yun, 2002). Slope length, slope steepness and slope aspect, which are three main features of a slope, affect not only the manner but also the amount of erosion. Slope steepness has a complex relationship with the process. It can increase the downward gravitational force of rainwater, thus increasing the runoff velocity, reducing the cumulative infiltration, and increasing the runoff sediment (Li et al., 2016). The ecological restoration measures have notable effects on soil erosion, which is greatly affected by human activities and has various degrees of influence on erosion (Wei & Jiao, 2021). These factors increase the complexity of the runoff and sediment yield mechanism and also increase the uncertainty of soil water erosion process. Compared to traditional soil erosion models, the machine learning model is able to predict erosion more promptly and efficiently (Liu et al., 2024c). Rahmati et al. (2017) employed multiple models to predict the soil erosion across lands for different purposes in Iran. Many Chinese researchers (e.g., Liu, Fan & Wang, 2024) have also studied the sediment yield of soil erosion in different regions using the machine learning model. Despite years of research, the Mu Us Sandy Land still lacks a machine learning-based model capable of predicting sediment yield from erosion and validating the prediction. Instead, remote sense imaging has been commonly used for simulation and prediction in this region, but its accuracy needs to be improved.

The present study aims to reveal the sediment yield from erosion in a quantitative approach with the machine learning model, according to the data collected by the soil and water conservation monitoring station in Miaochuan Basin from 2018 to 2023. Based on the basic features of runoff sediment, the researchers aim to achieve the following two objectives: 1. to select suitable machine learning models for the study area to simulate and validate the RD and SY under different ecological restoration measures; 2. to identify factors that directly or indirectly affect RD and runoff SY. The results provide theoretical guidance to the comprehensive prevention and control of soil erosion in the sand-covered pisha sandstone area.

Overview of the study area

The Kuye River Basin has the most serious sediment yield problem in the Yellow River (Zhao et al., 2019). The upper reaches of the Kuye River Basin are characterized by a landscape covered with sand dunes and shifting sands, while the lower reaches are located in a loess hilly and gully region. The natural ecosystem of the basin is fragile and the soil erosion is serious. It is also affected by human activities such as mineral exploitation and urban construction, which leads to structural and functional issues in the basin ecosystem, the obstruction of the hydrological cycle, and water environment and mine restoration problems (Zhao et al., 2019). The most typical sediment yield in the Kuye River Basin results from gravity erosion caused by rainstorm floods, and the maximum recorded sediment concentration is 1,700 kg/m³. Since 1980, the maximum sediment concentration of Wenjiachuan station is 1,420 kg m³, the maximum sediment discharge is 10.002 million tons, and the maximum sediment transport modulus is 1,157 tons km² (Wang et al., 2022). The study area was located on the right bank of the upper reaches of the four-level tributary of the Yellow River Basin, Ordos (109°31′30.97′′E, 39°39′2.89′′N) (Fig. 1). Situated at the northern edge of the Mu Us Sandy Land, the area is characterized by sand-covered hills and gullies. The topsoil of the region is loose and of no structure or a single-grain structure, with low water and nutrient retention capacity. The main types of soil include calcaric cambisol, aeolian sandy soil, pisha sandstone, and sand-covered pisha sandstone. Constructed in 2016, Hetongmiao Soil and Water Conservation Monitoring Station (see Fig. 2), Ejin Horo Banner, Ordos, was selected as the study area. The Miaochuan Basin located within the study area belongs to an arid and semiarid temperate continental climate. The average, maximum, and minimum annual rainfall are 358.2 mm, 642.7 mm, and 100.8 mm, respectively, and the average annual sunshine hours are 2,900. The effective accumulated temperature of ≥10 °C is 2,751.3 °C, and the average annual evaporation is 2,563 mm. The northwest wind dominates the area. The wind force is 5–8, the average annual wind speed is 3.6 m/s, and the maximum instantaneous wind speed is 24 m/s.

Materials & Methods

Three groups of standard runoff plots (5°, 10°, and 15°) were set up for runoff and sediment monitoring in the comprehensive monitoring station of soil and water conservation. The slope length of the runoff plot was 20 m, and the width was 5 m. Each group of plots was set up with Pinus sylvestris, Hippophae rhamnoides, artificial grassland, crops, natural grassland, and bare land. These plots of various surface forms basically represent the surface conditions of aeolian sandy soil, Pisha sandstone, and the hilly region of the Loess Plateau, and have a wide range of representative significance. The basic overview of runoff field ecological restoration measures (arbor forest land, shrub grassland, natural grassland, artificial grassland, cultivated-land, bare land) is shown in Table 1.

(1) Observation of rainfall characteristics.

The rainfall process was recorded by the weather station and observed every 5 min. The rainfall was accurate to 0.1 mm and the rainfall intensity was unit mm/min, combined with runoff and sediment yield (SY) on S under natural rainfall conditions in this area. For each rainfall event, the rainfall (P), rainfall duration (T), maximum 30-min rainfall intensity (I₃₀), and average rainfall intensity (I_avg) were selected, and the rainfall types in the study area were divided using the system clustering method. The characteristics of rainfall in the study area in 2018-2023 is shown in Table 2.

(2) Runoff sediment observation.

Using automatic water and sediment monitoring equipment (Beijing Tianhang Jiade, Beijing, China), the runoff volume was recorded once a minute, and the runoff volume was accurate to 0.001 L. The sediment yield was accurate to 0.001 kg/m³, supplemented by full-section samplers for manual field sampling. Runoff depth and sediment yield were calculated using the following formulas:

Runoff depth (RD) (mm) = Runoff volume (L)/Plot area (m²)

SY (t/hm²) = Total sediment (t)/Plot area (hm²).

(3) Vegetation observation.

The plant diversity, vegetation coverage, and height and quantity of 1 m × 1 m grass grid quadrats in each vegetation type plot were measured using the photographic method and visual estimation method from April to October, once every 15 days. The plant height, crown width, diameter at breast height, coverage, and spacing of trees and shrubs were measured by ruler from April to October. After the runoff of the plot, the canopy density of trees, coverage of shrub and grass crops, and ground coverage were measured.

(4) The machine learning model is employed.

The present study attempts to predict soil erosion with the neural network (NN), support vector regression (SVR), gradient boosting decision tree (GBDT), and K-nearest neighbors (KNN) and compare the performance of these models. The neural network is a representative of connectionist algorithms, the neuron being its smallest functioning unit. After the weighting of information X input into a neuron from the previous neuron, it is turned into information Y by a response function (e.g., sigmoid function). The foundation of the tree model is the decision tree. This method imitates the way humans make decisions, with the binary tree structure, such as yes/no and true/false (Breiman, 2001). Through data examination and evaluation, the scope of the answer is narrowed down. The support vector machine is a machine learning method based on the statistical learning theory. This model, characterized by high generalizability, strikes a balance between model complexity and learning ability with limited sample information.

(5) Statistical analysis.

The experimental data were collated and counted using Excel and plotted by Origin2018. SPASS26.0 was used for two-way analysis of variance (ANOVA), cluster analysis, and correlation analysis to study the connection between water and sediment under various ecological restoration measures. SPSS26.0 was used for path analysis, and the structural equation model was used for regression analysis. The machine learning model utilizes R software for simulation and validation.

Results and analysis

Runoff and sediment yield response to rainfall patterns under various ecological restoration measures

Across various rainfall types, there was a small variety of responses of RD or SY to S (Fig. 4). The RD and SY of bare land on each slope steepness were noticeably greater than other ecological restoration measures except for crop land. Overall, the steeper the slope, the greater the runoff depth and soil loss. On a 15-degree slope, which is considered steep, the erosive power of runoff increases. Accordingly, the effectiveness of soil and water conservation measures varies significantly. The Type III precipitation causes the shallowest runoff and lowest soil loss. Through correlation analysis of rainfall process, we found a correlation between RD, SY, T, P, I_avg, I_30,figS (Table 4). There was a noticeable (p <0.05) positive correlation between P, RD, and SY of various ecological restoration measures. The effects of T, I₃₀, and I_avg on runoff and sediment were not strongly correlated under various ecological restoration measures. S showed a noticeable (p < 0.05) or extremely noticeable (p < 0.01) positive correlation with RD in arbor forest land, natural grassland, cultivated land, and bare land. Table 5 reveals that there is a significant correlation between the RD and SY under various ecological restoration measures.

Table 4:

Correlation analysis of influencing factors of runoff and sediment yield under various control measures.

Control measures	Parameter	P	T	I30	Iavg	S
Arbor forest land	RD	.538*	0.169	0.144	0.212	0.999**
	SY	.750**	0.013	0.406	0.217	0.504*
Shrub grassland	RD	.762**	0.192	0.367	0.052	0.281
	SY	.611**	0.033	0.357	0.151	0.294
Natural grassland	RD	.510*	0.125	0.115	0.123	.576*
	SY	0.449*	0.017	0.243	0.118	0.315
Artificial grassland	RD	.575*	0.392	0.143	0.141	0.362
	SY	.652**	0.038	0.35	0.16	0.342
Cultivated-land	RD	0.494*	0.086	0.164	0.093	.717**
	SY	0.557*	0.018	0.263	0.163	0.426*
Bare land	RD	.530*	0.092	0.145	0.118	.498*
	SY	.588*	0.008	0.325	0.211	0.816**

DOI: 10.7717/peerj.20040/table-4

Notes:

Asterisks indicate whether there is a significant difference between factors. One asterisk indicates P < 0.05, and two asterisks indicate P < 0.01.

Table 5:

Correlation analysis of runoff and sediment under various control measures.

Parameter	RD
	Arbor forest land	Shrub grass land	Natural grassland	Artificial grassland	Cultivated- land	Bare land
SY	.841**	.788**	.787**	.722**	.894**	.666**

DOI: 10.7717/peerj.20040/table-5

Notes:

Asterisks indicate whether there is a significant difference between factors. Two asterisks indicate P < 0.01.

The structural equation model of influencing factors of RD and SY

Figure 5 is derived from the structural equation constructed based on factors affecting RD and the structural equation model (SEM). It was found that RD and soil erosion modulus increased as P, T, I_avg, I₃₀, and S increased. The influencing factors were positively correlated with RD and soil erosion modulus. Among them, P reached a noticeable level of P <0.01 for RD and soil erosion modulus. T and I_avg had a noticeable level of P <0.05 for RD. There is also a noticeable connection between RD and SY. In summary, the structural equation models reveal that T, I_avg, I₃₀ and S are factors directly influencing RD, which has direct implications on SY. Therefore, T, I_avg, I₃₀ and S are factors that indirectly affect SY.

Discussion

Conclusions

The present study is conducted on the data collected from long-term observations in standardized runoff plots where a variety of water and soil conservation and vegetation restoration have been taken. The results showed that the rainfall of the temple soil and water conservation monitoring station was mainly concentrated from July to September. Among them, the rainfall events with short duration, medium rainfall, and medium rainfall intensity were the most frequent. All the ecological restoration measures pointed out that the rainfall events with medium duration, heavy rainfall, and heavy rainfall intensity had the most serious runoff and sediment.

Compared with the control plot (bare land), ecological restoration measures can noticeably reduce runoff and SY. The influence of ecological restoration measures on RD and SY varied with ecological restoration measures. Among them, arbor forest and shrub grassland were the best ecological restoration measures for controlling soil loss because of their canopy structure. Tree- and shrub-forests are effective in preventing and control soil erosion. Using machine learning models, the researchers found that GBDT performed best in simulating soil erosion in the study area. The structural equation model showed that there is a significant correlation between RD and SEM. T, I_avg, I₃₀ and S are factors that directly affect SY. Since RD directly impacts SY, it can be said that T, I_avg, I₃₀ and S indirectly influence SY. Our research findings can provide theoretical guidance and technical support for the prevention and control of soil erosion under different slope management measures in the sand-covered soft rock area of the middle reaches of the Yellow River, thereby contributing substantially to the ecological restoration in the Yellow River Basin.

Supplemental Information