论文信息

Xinbin Li / Yilong Zhu / Chengxin Kang 等 9 人

Scientific Reports

论文解读

AI 辅助整理
方法
RUSLE模型估算2000—2024年侵蚀模数,构建XGBoost-SHAP可解
发现
地质基质是侵蚀空间分异最密切的背景因子,软岩与黄土区为持续高风险区,2020年极端降雨致侵蚀反弹。
机会点
可引入跨区域迁移验证与地质分层建模,弥补RUSLE结构重叠导致的解释偏差,提升极端气候下侵蚀预测稳健性。

论文摘要

量化土壤侵蚀的时空动态并阐明其相关驱动因素,是制定针对性水土保持策略的重要前提。本研究以黄河中游主要粗泥沙来源区窟野河流域为研究对象,采用RUSLE模型估算2000—2024年土壤侵蚀模数,分析侵蚀的时空演变特征与地质分异规律。构建了可解释的XGBoost-SHAP机器学习框架,纳入气候、地形、植被、人为活动和地质基底等因素,定量解析各因素对RUSLE估算侵蚀空间格局的贡献。设计七组控制实验,检验RUSLE内部因子与模型自变量之间的结构重叠。结果表明:(1) 土壤侵蚀强度总体呈改善趋势,但空间格局保持稳定,并表现出显著岩性梯度。风成黄土、泥质岩及碎屑岩夹泥质岩单元被识别为持续性高侵蚀风险区。(2) 地质基底是与侵蚀空间分异关系最为密切的背景因素。单因子敏感性空间分布显示,K因子高敏感区与泥质岩和风成沙分布一致,LS因子高敏感区与黄土丘陵沟壑区陡坡地带吻合,C因子高敏感区与中下游低植被覆盖区重叠。三因子的空间叠加与高风险单元持续性高侵蚀风险密切相关。2020年极端降雨下全流域侵蚀反弹,凸显了软弱基底对极端气候的敏感性。(3) 在RUSLE框架内,对应C因子和LS因子的NDVI与坡度对侵蚀模数空间格局的贡献最大。然而控制实验表明,该结果部分源于RUSLE方程的数学结构关系,模型性能主要反映其在随机留出验证下重构RUSLE输出的能力,而非跨区域预测迁移能力。外部环境变量(PET、Distance

全文

Scientific Reports

ARTICLE | Article in Press
https://doi.org/10.1038/s41598-026-79921-0
This Article in Press is shared early to give you faster access to new research. It is citable and carries a permanent DOI.
The final edited version will replace it automatically.
Scientific Reports
Scientific Reports
Geological substrate and multi-factor controls on
RUSLE-estimated water erosion in the Kuye River Basin
over 25 years
Xinbin Li · Yilong Zhu · Chengxin Kang · Junsheng Xi · Hao Zhao · Duoxun Xu ·
Zhiqiang Yin · Chaohui Han · Jianghua Zhang
Received: 6 August 2026 / Accepted: 4 November 2026
© The Author(s) 2026
Abstract
Quantifying the spatiotemporal dynamics of soil erosion and clarifying its associated driving factors are essen -
tial prerequisites for developing targeted soil and water conservation strategies. This study focused on the
Kuye River Basin, a primary coarse sediment source in the middle Yellow River. The RUSLE model was
used to estimate the soil erosion modulus from 2000 to 2024, and the spatiotemporal evolution and geological
differentiation of erosion were analyzed. An interpretable XGBoost-SHAP machine learning framework was
constructed, incorporating climatic, topographic, vegetation, anthropogenic, and geological substrate factors,
to quantitatively disentangle the contributions of each factor to the RUSLE-estimated erosion spatial pattern.
Seven groups of controlled experiments were designed to examine the structural overlap between RUSLE
internal factors and model independent variables. The results show that: (1) The overall soil erosion intensity
exhibited an improving trend, but the spatial pattern remained stable with a pronounced lithological gradient.
Aeolian loess, argillaceous rock, and clastic rock interbedded with argillaceous rock units were identified as
persistent high erosion risk zones. (2) Geological substrate was the background factor most closely related
to the spatial differentiation of erosion. The spatial distribution of single-factor sensitivity showed that high-
sensitivity areas of the K factor coincided with argillaceous rock and aeolian sand, high-sensitivity areas of the
LS factor were consistent with steep-slope zones in the loess hilly-gully region, and high-sensitivity areas of
the C factor overlapped with low-vegetation-cover areas in the middle and lower reaches. The spatial superpo -
sition of these three factors is closely associated with the persistently high erosion risk in the high-risk units.
The basin-wide erosion rebound in 2020 under extreme rainfall highlights the sensitivity of soft substrates to
extreme climate. (3) Within the RUSLE framework, NDVI and slope, corresponding to the C and LS factors,
contributed most to the spatial pattern of erosion modulus. However, controlled experiments indicated that this
result partly arises from the mathematical structural relationship of the RUSLE equation, and the model perfor -
mance primarily reflects its capacity to reconstruct RUSLE outputs under random holdout validation rather than
cross-regional predictive transferability. External environmental variables (PET, Distance to roads, Distance to
water, Aspect, Soil thickness, and Lithology) showed limited independent explanatory power for RUSLE out -
puts (M5 R2 = – 0.011). Therefore, the SHAP results should be interpreted as the “covariate importance of the

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
RUSLE-estimated erosion pattern” rather than the “causal mechanisms of actual erosion processes.” It is rec -
ommended that geological unit differentiation be incorporated as a core consideration in future soil and water
conservation planning, and that a precision governance system based on zoning of rock-soil erosion resistance
be established. Future studies should integrate independent observational data (e.g., check dam sedimentation,
137Cs tracing, and sediment discharge at hydrological stations) to distinguish RUSLE structural associations
from actual erosion drivers.
Keywords Soil erosion · RUSLE · XGBoost-SHAP · Geological substrate · Covariate importance · Kuye
River Basin · The middle Yellow Rive
Introduction
Soil erosion is among the most pressing eco-environmental challenges worldwide 1. It not only causes land deg-
radation such as soil fertility decline, but also triggers a series of cascading effects including downstream river
siltation and aggravated drought and flood disasters, severely threatening regional ecological security and high-
quality development2,3. The coarse-sediment producing area in the middle reaches of the Yellow River is one of
the regions with the most intense soil erosion and the most fragile ecological environment in China, and also the
core source of coarse sediment for the Yellow River main stem. Soil and water loss control in this area is a core
task of the national strategy for Ecological Protection and High-Quality Development of the Yellow River Basin4,
and directly affects the ecological security of the entire basin 5,6. Since the founding of the People’s Republic of
China, the state has implemented a series of comprehensive soil and water conservation projects in this region 7,
which have significantly improved local ecological conditions 8 and effectively reduced sediment loads entering
the Yellow River9. However, under the combined influence of complex natural background conditions and inten-
sifying climate change, soil erosion remains at high risk in some areas, and even rebounds in years with extreme
rainfall. Particularly in zones with weak lithology, steep terrain and low vegetation coverage, erosion processes
respond drastically to extreme rainfall events, exposing the vulnerability and limitations of existing control mea-
sures in coping with high-intensity disturbances 10,11. Previous studies have mostly focused on the impacts of
vegetation cover, topography and geomorphology, and human activities on erosion. By contrast, insufficient
attention has been paid to the contribution and underlying mechanism of geological substrate as the fundamental
background for the initiation and development of erosion, making it difficult to support precision soil and water
conservation planning based on watershed natural endowments12.
The Kuye River Basin lies in the transitional zone between the Mu Us Sandy Land and the loess hilly and gully
region in the middle reaches of the Yellow River. Featuring remarkable geological substrate heterogeneity and
fragmented topography, the basin shows a prominent controlling effect of stratigraphic lithology on soil erosion
processes13. With pronounced climatic characteristics of alternating cold and warm periods, concentrated rainfall
and dry-wet cycles, superimposed by human disturbances such as intensive exploitation of coal resources, this
basin is a typical tributary in the coarse-sediment producing area of the middle Yellow River 14, and also serves
as a representative watershed for investigating the spatiotemporal differentiation patterns of soil erosion and its
driving mechanisms15. In recent years, a large number of studies have been carried out in the Kuye River Basin,
involving runoff variation and attribution16, impacts of coal mining on groundwater17, future runoff responses to
coal mining expansion and climate change18, responses of vegetation cover to climate change and human activi-
ties19, landscape ecological risk and human activity intensity 20,21, evolution of land use and ecosystem service
value22, construction and optimization of watershed ecological networks23, and remote sensing ecological index
assessment24. These studies provide an important basis for understanding the ecological-hydrological processes
of the basin, but most of them focus on a single process or specific element, and the role of geological substrate
in the erosion driving system still lacks systematic quantification. Chen (2015) conducted an early assessment
of erosion sensitivity in the basin using the RUSLE model and identified high-risk erosion zones 25. He (2025)

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
further predicted the evolution trend of erosion by integrating the CA-Markov model, and confirmed the long-
term erosion reduction effect of the Grain for Green Project26. In addition, Kou et al. (2021) revealed the charac-
teristics of rill development and its change rate under constant rainfall intensity through field experiments with
artificially simulated rainfall 27. In 2024, they further used unmanned aerial vehicle (UA V) photogrammetry to
assess the erosion characteristics and hydrological effects of loess terraced slopes28, and employed SBAS-InSAR
technology to analyze gully erosion and deposition patterns in loess tablelands 29. These studies provide multi-
scale observational evidence for understanding erosion processe. Overall, however, existing studies have mostly
focused on the macroscopic estimation of water erosion, with insufficient attention to the erosion differences of
special rock-soil media widely distributed in the basin, including Pisha sandstone, aeolian sand and loess. The
integration of erosion estimation with geological substrate properties remains to be further deepened.
Soil erosion models serve as critical tools for the quantitative assessment of soil erosion intensity 30, and are
mainly categorized into physically based models and empirical-statistical models 31. Physically based models,
such as WEPP and LISEM, are constructed on the physical processes of erosion and sediment yield with clear
mechanisms. However, they demand numerous parameters and high data availability, which imposes consid -
erable limitations on their application and promotion in large-scale watersheds 5. Among empirical-statistical
models, the Universal Soil Loss Equation (USLE) and its revised version RUSLE have become the most widely
used erosion estimation tools worldwide, owing to their simple structure, readily available parameters and strong
regional adaptability32–34. On the basis of quantitative erosion estimation, identifying driving factors and their
action mechanisms is a key prerequisite for achieving precision soil and water loss control. Early studies on driv-
ing mechanisms mostly adopted traditional statistical methods such as correlation analysis and multiple regres -
sion, which fail to effectively characterize the nonlinear relationships between erosion and various driving factors
as well as the interaction effects among multiple factors 35. The introduction of the Geodetector model provides
a powerful tool for identifying the driving power of factors and quantifying their interactions, and has been
gradually applied to research on soil erosion driving mechanisms. For instance, Liu et al. (2025) found in their
study on the Shanxi section of the Yellow River Basin that vegetation coverage was the dominant factor of soil
erosion, and the explanatory power of the interaction between vegetation and slope was significantly higher than
that of any single factor 36. Nevertheless, the Geodetector model requires discretization of continuous variables,
which tends to cause loss of original information. Furthermore, it is difficult to quantify the marginal effects of
individual factors on erosion and their nonlinear action patterns 37. In recent years, machine learning algorithms
have provided new approaches for the analysis of erosion driving mechanisms, given their outstanding capabil -
ity in processing nonlinear and high-dimensional data. Among them, the eXtreme Gradient Boosting (XGBoost)
algorithm is emerging as an important tool for driving factor analysis in the eco-environmental field, due to its
high prediction accuracy, strong generalization ability and effective overfitting control 38. However, traditional
machine learning models generally have a “black-box” nature, making it difficult to reveal the direction of con -
tribution and mode of action of individual factors. The SHAP method, based on Shapley values from game
theory, enables interpretability analysis of machine learning models and quantitatively characterizes the marginal
contribution of each feature to prediction results, providing a new technical pathway for the refined analysis of
driving mechanisms37. At present, the XGBoost-SHAP framework has been preliminarily applied in fields such
as vegetation dynamics and ecological security assessment 39,40. Nevertheless, its application in research on soil
erosion driving mechanisms remains limited. In particular, studies that incorporate geological lithology factors
into the driving system and quantitatively analyze the regulatory role of lithological substrates are even more
scarce. It should be noted that when the prediction target of a machine learning model is itself derived from an
empirical model such as RUSLE, structural overlap may exist between the independent variables and the depen-
dent variable, causing the model’s importance ranking to partly reflect formula construction relationships rather
than independent driving mechanisms. This study will systematically examine this issue and, on this basis, limit
the interpretation scope of the conclusions.
Against this background, this study takes the Kuye River Basin as the study area. Based on multi-source spa -
tiotemporal data covering meteorology, land use, vegetation, topography, soil properties and geological lithology,

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
we adopt the RUSLE model to estimate the soil erosion modulus of the basin from 2000 to 2024, and systemati-
cally analyze the spatiotemporal dynamic evolution of soil erosion and the geological differentiation characteris-
tics of erosion risk over the past 25 years. On this basis, we construct an interpretable XGBoost-SHAP machine
learning model that incorporates geological substrate factors including lithology and soil thickness. From five
dimensions of climate, topography, vegetation, human activities and geological substrate, we quantitatively elu-
cidate the contribution intensity, action direction, nonlinear patterns and interaction effects of each driving factor.
This study aims to reveal the underlying regulatory mechanism of the geological lithological substrate on erosion
intensity, as well as its synergistic or restrictive relationships with vegetation and topography. The results can
provide a scientific basis for precise soil and water loss prevention and ecological restoration guided by the dif -
ferentiation characteristics of geological units in the coarse-sediment producing area of the middle Yellow River.
Data and methods
Study area
The Kuye River Basin spans Ejin Horo Banner of Ordos City, Inner Mongolia Autonomous Region, and Shenmu
City of Shaanxi Province. As a first-order tributary on the right bank of the Jin-Shaan Gorge in the middle reaches
of the Yellow River, it has a main stream length of 242 km and a total drainage area of 8706 km² (Fig. 1). Located
in the ecotone between the Mu Us Sandy Land and the hilly and gully region of the Loess Plateau, the basin fea-
tures a terrain high in the northwest and low in the southeast, with elevation ranging from 736 m to 1363 m. It has
Fig. 1 Location of the study area. The basemap is the World Terrain Base provided by Esri.(Sources: Esri, USGS, NOAA).
National Catalogue Service for Geographic.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
an arid to semi-arid temperate continental monsoon climate, with a mean annual precipitation of approximately
410 mm. Rainfall is highly concentrated from June to September, mostly in the form of rainstorms. Potential
evaporation far exceeds precipitation, resulting in extremely scarce water resources.
Information of China, National County-level Administrative Boundaries Shapefile Data ( h t t p s : / / w w w . w e b m
a p . c n / m a i n . d o ? m e t h o d = i n d e x ) ; Geospatial Data Cloud, 30-m resolution digital elevation data ( h t t p s : / / w w w . g s
c l o u d . c n / ) . The map was created using ArcGIS Pro v.3.7.0 software by Esri (Environmental Systems Research
Institute, Redlands, CA, USA; https://www.esri.com/).
Geotectonically, the basin belongs to the Yishan Slope in the northeastern Ordos Basin. Controlled by geo -
logical structure and sedimentary environment, the lithostratigraphy of the basin shows remarkable spatial dif -
ferentiation. In the upper reaches, the landscape is dominated by sand-covered Pisha sandstone hills mantled
by modern aeolian sand, where Cretaceous sandstone and mudstone of the Yijinhuoluo Formation and Luohe
Formation are mainly developed. The middle and lower reaches gradually transition into sand-covered loess hills
and typical loess hilly landforms. Jurassic sandstone and mudstone of the Yan’an, Zhiluo and Anding forma -
tions are widely exposed, overlain by discontinuous Malan Loess and Lishi Loess. The rock and soil masses are
generally loose with low cementation and weak erosion resistance (Fig. 2a). Soil thickness in the basin ranges
from 138 to 233 cm. Spatially, it is thinner in the north (138 ~ 180 cm) and thicker in the central and southern
parts (193 ~ 233 cm), consistent with the spatial differences in aeolian sand cover and loess deposition (Fig. 2b).
Slopes in the basin are dominated by gentle slopes of < 15°, but steep slopes of 15°~35°, locally exceeding 35°,
are developed in the loess hilly-gully region and on both sides of river valleys, providing favorable topographic
conditions for erosion and sediment yield (Fig. 2c). Land use is dominated by grassland; forestland and shrubland
are mainly distributed in the upper reaches and on gully slopes; built-up land is concentrated in belts or patches
along river valleys, towns and mining areas; water bodies comprise river channels and reservoirs; and unused
land is scattered (Fig. 2d). Meanwhile, the basin is located in the core area of the Shenfu-Dongsheng Coalfield.
High-intensity coal mining has led to groundwater level decline, intensified surface disturbance, accelerated
bedrock weathering, and local vegetation degradation, further exacerbating erosion risk. Together, these multiple
factors make the Kuye River Basin a typical tributary with high-intensity coarse sediment yield in the middle
reaches of the Yellow River.
Data sources and processing
The meteorological data used in this study were derived from the Basic Meteorological Element Observation
Dataset of China’s National Surface Meteorological Stations ( h t t p s : / / d a t a . c m a . c n / d a t a / c d c d e t a i l / d a t a C o d e / A . 0 0 1
2 . 0 0 0 1 . h t m l ) . The land use data were obtained from the China Land Cover Dataset (CLCD, h t t p s : / / d o i . o r g / 1 0 . 5 2 8
1 / z e n o d o . 1 8 1 8 0 1 8 4 ) 4 1 . The potential evapotranspiration data came from the 1 km Monthly Potential Evapotrans-
piration Dataset of China ( h t t p s : / / w w w . g e o d a t a . c n / d a t a / d a t a d e t a i l s . h t m l ? d a t a g u i d = 3 4 5 9 5 2 7 4 9 3 9 6 2 0 & d o c I d = 2 1
5 2 ) 4 2 . The Normalized Difference Vegetation Index (NDVI) data were sourced from the 30 m NDVI Dataset of
China’s Loess Plateau (http://loess.geodata.cn)43. The DEM data were extracted from the 30 m Aster GDEM Dig-
ital Elevation Model hosted on the Geospatial Data Cloud (https://www.gscloud.cn/). The soil property data were
acquired from the High-Resolution National Soil Grid Basic Attribute Dataset of China ( https://www.geodata.
cn/data/datadetails.html?dataguid=36810085119113&docId=0)44. The stratigraphic l i t h o l o g y data were retrieved
from the spatial database of the public-version 1:200,000 national digital geological map ( http://dcc.cgs.gov.
cn)45. All spatial datasets were processed with projection transformation, resampled to a uniform resolution of 30
m, and clipped to the boundary of the Kuye River Basin to ensure the consistency of model inputs. It should be
noted that resampling data of different resolutions to 30 m is only for spatial alignment and does not add effective
spatial information to the original data. The analysis results of factors such as lithology and soil thickness should
be interpreted at the sub-basin or regional scale, rather than at the 30 m plot scale. All intermediate datasets and
results generated during this study are available from the corresponding author upon reasonable request. Detailed
information on all data sources is listed in Table 1.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
Fig. 2 Map of surficial substrate types (a), spatial distribution of soil thickness (b), slope classification map (c), and land use
map of 2024 (d) in the Kuye River Basin. The map was created using ArcGIS Pro v.3.7.0 software by Esri (Environmental
Systems Research Institute, Redlands, CA, USA; https://www.esri.com/).

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
Soil erosion estimation methodology
Soil erosion was quantitatively estimated using the Revised Universal Soil Loss Equation (RUSLE) on the Arc -
GIS platform, as expressed by the following equation:
A = R × K × LS × C × P (1)
Where A represent soil erosion modulus per unit area, with a unit of t·hm− 2·a− 1; R represent rainfall erosivity factor,
with a unit of MJ·mm·hm− 2·h− 1·a− 1; K represent soil erodibility factor, with a unit of t·hm2·h·hm− 2·MJ− 1·mm− 1;
LS represent slope length-steepness factor (dimensionless); C represent vegetation cover and management fac -
tor (dimensionless); P represent support practice factor (dimensionless). Erosion intensity grades were classified
in accordance with the Standards for Classification and Gradation of Soil Erosion (SL 190–2007) issued by the
Ministry of Water Resources of the People’s Republic of China (2008)46.
(1) Rainfall erosivity factor (R). Based on daily precipitation data from 34 meteorological stations in and adja -
cent to the study area, the R factor is calculated using the monthly-scale model proposed by Wischmeier and
Smith (1978)32. The formula is as follows:

R =
12
∑
𝑖𝑖𝑖1
1.735 × 10
[(1.5×lg
𝑝𝑝2𝑖𝑖
𝐴𝐴𝐴𝐴 )−0.8188] (2)
where AP is the annual precipitation (mm), and Pi is the precipitation of the i-th month (mm).
The detailed location information and spatial distribution of the 36 meteorological stations used in this study
are provided in Table S1 and Figure S1, respectively. The rainfall data processing methods, spatial interpolation
Table 1 Data sources and processing details of the study.
Data type Dataset name Primary purpose Source Spatial
resolution
Time
span
Precipitation China National Surface
Meteorological Station
Basic Meteorological Ele-
ment Data Product
Calculation of rainfall
erosivity factor (R)
National Meteorological Information Cen-
ter; h t t p s : / / d a t a . c m a . c n / d a t a / c d c d e t a i l / d a t a
C o d e / A . 0 0 1 2 . 0 0 0 1 . h t m l
Station-based 2000–
2024
Land cover China Land Cover Dataset
(CLCD)41
Extraction of support
practice factor (P)
https://doi.org/10.5281/zenodo.18180184 30 m 2000–
2024
Potential
evapotranspiration
1 km Monthly Potential
Evapotranspiration Dataset
for China42
Extraction of poten-
tial evapotranspira-
tion (PET)
National Earth System Science Data
Center; h t t p s : / / w w w . g e o d a t a . c n / d a t a / d a t a d
e t a i l s . h t m l ? d a t a g u i d = 3 4 5 9 5 2 7 4 9 3 9 6 2 0 & d
o c I d = 2 1 5 2
1 km 2000–
2024
NDVI/Vegetation
coverage
30 m NDVI/Vegetation
Coverage Dataset of the
Loess Plateau43
Calculation of vegeta-
tion cover factor (C)
Loess Plateau Science Data Center; h t t p : / / l
o e s s . g e o d a t a . c n
30 m 2000–
2024
Topography ASTER GDEM 30 m
Digital Elevation Model
Watershed boundary
delineation; calcula-
tion of slope length-
steepness factor (LS)
Geospatial Data Cloud (GSCloud); h t t p s : / /
w w w . g s c l o u d . c n /
30 m —
Soil properties High-resolution National
Soil Information Grid
Basic Attribute Dataset of
China44
Calculation of soil
erodibility factor (K);
Extract soil thickness
information
National Earth System Science Data
Center; h t t p s : / / w w w . g e o d a t a . c n / d a t a / d a t a
d e t a i l s . h t m l ? d a t a g u i d = 3 6 8 1 0 0 8 5 1 1 9 1 1 3 &
d o c I d = 0
90 m 2010–
2018
Stratigraphic
lithology
National 1:200,000
Digital Geological Map
(Public Version) Spatial
Database45
Division of surface
geological units
China Geological Survey Data Center; h t t p
: / / d c c . c g s . g o v . c n
1:200,000 —

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
methods, results, and the interpolation reliability verification process are presented in Sect. 2 of the Supplemen -
tary Material, Figure S2, and Table S2.
(2) Soil erodibility factor (K). The K factor is calculated using the algorithm from the EPIC (Erosion Productiv-
ity Impact Calculator) model47. The formula is as follows:

KEPIC =0.1317 × {0.2 + 0.3 × exp [−0.0256 × SAN × (1− SIL100 )]}
× ( SILCLA + SIL )
0.3
× [1.0 − 0.25 × CC + exp (3.72 − 2.95 × C) ]
× [1.0 − 0.7 × SN1SN1+ exp (−5.51 + 22.9 × SN1) ]
(3)
where CLA, SIL and SAN represent the contents of clay, silt and sand (%), respectively; C denotes the soil
organic carbon content (%); and SN1 = 1 − SAN/100. The final K value is derived from the conversion formula:
K=( − 0.01383 + 0.51575KEPIC)×0 .1317. The estimation method for the soil erodibility factor K used in this study has
been widely validated in the Loess Plateau region4,48.
(3) Topographic factor (LS). Based on the DEM data, the slope length factor ( L) and slope steepness factor ( S)
are extracted using the method proposed by Liu et al.33,49:
L=( 𝜆𝜆22.13 )m (4)

m=
⎧{{
⎨{{⎩
0.2, 𝜃𝜃 = 0.5
∘
0.3, 0.5
∘
<𝜃𝜃 = 1.5
∘
0.4, 1.5
∘
<𝜃𝜃 = 2.5
∘
0.5, 𝜃𝜃𝜃 2.5
∘
(5)

S=
⎧{
⎨{⎩
10.8sin𝜃𝜃 + 0.03, 𝜃𝜃𝜃 5
∘
16.8sin𝜃𝜃𝜃 00.5, 5
∘
= 𝜃𝜃 = 10
∘
21.9sin𝜃𝜃𝜃 00.96, 𝜃𝜃𝜃 10
∘
(6)
where λ denotes the slope length (m), m is the slope length exponent, and θ represents the slope gradient (°).
(4) Vegetation cover and management factor ( C). Fractional vegetation coverage is first calculated using the
dimidiate pixel model:

FVC= NDVI-NDVImin
NDVImax+NDVImin (7)
The C value is subsequently calculated using the formula proposed by Cai et al.50:

C= {
1, 0 = FVC = 0.1
0.650800.3436lg(FVC), 0.1 < FVC = 0.783
0, FVC > 0.783
(8)
where FVC denotes fractional vegetation coverage; NDVI is the normalized difference vegetation index; NDVImin
represents the minimum NDVI value in the study area (corresponding to pure bare soil pixels); and NDVImax rep-
resents the maximum NDVI value in the study area (corresponding to pure vegetation pixels).
(5) Support practice factor (P). Referring to the assignment methods applied in the middle reaches of the Yellow
River and the Kuye River Basin by Zhang et al. 4 and He et al.26, P values are determined according to land
use types: 0.30 for cropland, 1 for forest land, shrubland, grassland and unused land, and 0 for water bodies
and construction land. However, this direct assignment method fails to account for the dynamic changes of

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
engineering measures such as terraces, check dams, mining disturbances, and roads, which may exert a cer -
tain influence on the magnitude of erosion trends. Future studies should conduct dedicated sensitivity tests
of the P factor by incorporating disturbance factors such as mining activities and soil and water conservation
engineering measures.
Methods for analyzing the associative driving structure
In this study, a coupled XGBoost-SHAP model was employed to decipher the associative driving structure under-
lying variations in erosion intensity. eXtreme Gradient Boosting (XGBoost) is an ensemble learning algorithm
built on the gradient boosting framework. With the incorporation of regularization terms and parallel comput -
ing capability, it can efficiently handle high-dimensional data and capture nonlinear associations among driving
factors38.
The annual soil erosion modulus at pixel scale is set as the dependent variable. Ten indicators are selected as
independent variables, namely annual precipitation, potential evapotranspiration, slope gradient, slope aspect,
NDVI, land use type, distance to roads, distance to drainage networks, surface substrate type and soil thickness,
which fall into five categories of driving factors: climate, topography, vegetation, human activities and geologi -
cal substrate. To evaluate model performance, the dataset was randomly divided into an 80% training set and a
20% test set. Five-fold cross-validation was adopted, in which the training set was randomly partitioned into five
non-overlapping subsets; four subsets were used for model training and the remaining one for model validation.
Each time a different subset was selected as the validation set, and this process was repeated five times to traverse
the entire training set, thereby achieving cross-validation. It should be noted that, given the spatial autocorrela -
tion inherent in pixel-based data, random splitting may place spatially adjacent samples in both the training and
validation sets, which could potentially lead to optimistic performance estimates. The validation scores from the
five rounds were averaged as the final evaluation result of the model’s reconstruction performance. This valida -
tion primarily assesses the model’s capacity to reproduce RUSLE outputs within the study domain, rather than
establishing predictive transferability across different regions or years. Key parameters were optimized using a
grid search algorithm. Multiple candidate values were pre-specified for the hyperparameters of the target model,
all combinations of parameters were traversed, and the effect of each parameter combination was tested with
cross-validation. Finally, the optimal hyperparameter combination was selected, with the maximum tree depth set
to 5, the learning rate set to 0.1, and the number of iterations set to 100. The model’s fitting degree and stability
were evaluated by calculating R², root mean square error (RMSE), and mean absolute error (MAE) on the test set.
To enhance the interpretability of the model, the SHAP (SHapley Additive exPlanations) method is adopted.
Grounded in cooperative game theory, SHAP decomposes the model’s predicted values into a linear sum of
contribution values from each feature. It enables both global feature importance ranking and local interpretation
of individual samples simultaneously, and quantitatively characterizes the marginal contribution of each feature
to the prediction results 51. Similar interpretable machine learning frameworks have been successfully applied
in landslide susceptibility assessment 52,53 and ecological risk assessment 21, validating the effectiveness of this
method in driving mechanism analysis within complex geographical environments.
The calculation formula for the SHAP value of feature i is expressed as:

∅ i = ∑
S∪M{i}
|S|!(|N| − |S| − 1)!
|N|! [𝑣𝑣(S − {i})−𝑣𝑣 (S)] (9)
where N is the full set of all features, S denotes any subset excluding feature i, and 𝜈𝜈 () represents the predicted
output of the model given a specific feature subset. 𝑣𝑣 (S ∪ {i}) is the predicted value obtained after adding feature
i to subset S. ∅ i denotes the SHAP value of feature i: a positive value indicates that the feature promotes soil
erosion, whereas a negative value indicates an inhibitory effect, and the magnitude of its absolute value reflects
the strength of the contribution. In this study, the mean absolute SHAP value is adopted to quantify the overall

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
contribution intensity of each driving factor. In addition, SHAP dependence plots are used to analyze the marginal
effects of key factors on soil erosion across different value ranges. It should be emphasized that in this study, the
XGBoost model was employed to analyze the statistical associations between RUSLE estimates and various fac-
tors, rather than to establish causal mechanisms. SHAP values reflect the marginal contributions of each factor to
model predictions and cannot be directly interpreted as causal drivers of actual erosion processes.
Collinearity diagnosis and circular reasoning test
Collinearity diagnosis
The Spearman rank correlation coefficient was used to evaluate the monotonic correlations among variables. This
method is robust to non-normal distributions and outliers, and is suitable for erosion data containing categorical
variables and a large number of zero values. The correlation coefficient matrix between the dependent variable
and each independent variable, as well as among independent variables, was calculated, and significance tests
were performed ( p < 0.05). VIF was used to detect multicollinearity among independent variables. For the i-th
independent variable, VIF is defined as:

𝑉𝑉𝑉𝑉𝑉𝑉 𝑖𝑖 = 11−𝑅𝑅 2𝑖𝑖
where 𝑅𝑅2 is the coefficient of determination obtained by performing Ordinary Least Squares regression (OLS) of
the i-th independent variable on the remaining independent variables. VIF > 10 indicates severe collinearity, and
VIF > 5 indicates moderate collinearity. In this study, VIF was calculated using the closed-form solution of the
correlation matrix inversion, i.e., VIF = diag(R−1), where R is the Spearman correlation matrix among independent
variables. Categorical variables were one-hot encoded (drop_first = True) before entering the calculation.
Controlled experiment design for model factors
To quantify the structural overlap between RUSLE internal factors and XGBoost independent variables, seven
groups of controlled experiments were designed in this study. The detailed variable set design is shown in Table
S5. M0 is the full-variable baseline model; M1 excludes NDVI (a proxy for the C factor); M2 excludes Slope (a
proxy for the LS factor); M3 excludes both NDVI and Slope; M4 excludes all RUSLE direct factors (Precipita -
tion, NDVI, Slope, and LULC); M5 retains only non-RUSLE variables (Potential evapotranspiration, Distance to
roads, Distance to water, Aspect, Soil thickness, and Lithology); and M6 randomly permutes NDVI and Slope to
test whether their importance depends on spatial structure.
Results
Spatiotemporal evolution pattern of soil erosion
Estimation of the soil erosion modulus in the Kuye River Basin from 2000 to 2024 derived from the RUSLE
model shows that soil erosion intensity exhibits distinct spatiotemporal differentiation characteristics (Fig. 3).
Spatially, the erosion pattern across the upper, middle and lower reaches is highly consistent with the geomor -
phic-lithological zonation of the basin. The sand-covered Pisha sandstone hilly area in the upper reaches is domi-
nated by moderate erosion, with localized severe erosion. The sand-covered loess hilly area in the middle reaches
features markedly higher erosion intensity, where very severe and extreme erosion occurs in contiguous patches.
The erosion intensity decreases to some extent in the loess hilly area of the lower reaches, but remains dominated
by moderate and severe erosion. This spatial pattern remained generally stable over the study period.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
Temporally, the overall spatial pattern of soil erosion intensity in the basin remained largely stable from 2000
to 2024. From 2000 to 2015, the basin witnessed a phased decline in overall erosion intensity: the areas of very
slight and slight erosion expanded continuously, while the area of extreme erosion shrank significantly, indicat -
ing that soil and water conservation measures have achieved favorable erosion reduction outcomes. However,
a prominent inflection point of “erosion rebound” emerged in 2020. All erosion intensity grades shifted toward
higher levels, and the area of extreme erosion rebounded markedly. By 2024, erosion intensity had decreased
somewhat relative to 2020, yet the proportion of areas with erosion above moderate grade had not recovered to
the 2015 level, suggesting that extreme events exert persistent impacts on the regional erosion pattern (Fig. 4).
Erosion risk identification based on geological units
Statistics of soil erosion modulus across different geological units reveal that rock and soil types exert hierarchi-
cal control over erosion intensity (Fig. 5). The aeolian loess unit records the highest multi-year average soil ero-
sion modulus at 11595.11 t·km⁻²·a⁻¹, classified as an extremely high-risk unit. It is followed by the argillaceous
rock unit (8290.15 t·km⁻²·a⁻¹), medium clastic rock interbedded with argillaceous rock unit (8028.43 t·km⁻²·a⁻¹)
and fine clastic rock interbedded with argillaceous rock unit (7319.13 t·km⁻²·a⁻¹), all categorized as high-risk
bedrock units. The coarse clastic rock and alluvial-proluvial soil units, with erosion moduli ranging from 3000
to 4600 t·km⁻²·a⁻¹, fall into the moderate risk category. The lacustrine soil unit has the lowest erosion modulus
(1247.32 t·km⁻²·a⁻¹) and the strongest erosion resistance. Temporally, the erosion modulus of most units gener -
ally showed a downward trend from 2000 to 2015, but all exhibited a distinct synchronous peak in 2020. Among
them, the aeolian loess and argillaceous rock units showed the most remarkable rebound magnitude.
Fig. 3 Spatial distribution of soil
erosion intensity grades in the Kuye
River Basin from 2000 to 2024. The
map was created using ArcGIS Pro
v.3.7.0 software by Esri (Environ-
mental Systems Research Institute,
Redlands, CA, USA; h t t p s : / / w w w . e
s r i . c o m / ) .

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
The compositional structure of erosion intensity by area share further reveals the internal erosion heterogeneity
within geological units (Fig. 6). For units with weak substrates such as aeolian soil and argillaceous rock, the pro-
portion of very slight erosion area consistently stays at 30%~55% year-round, while the share of areas with mod-
erate and higher erosion remains above 40% over the long term. Among all unit types, these weak substrate units
have the highest proportion of extreme erosion area. In contrast, units such as coarse clastic rock and lacustrine
soil maintain a stable share of very slight erosion at 50%~85%, with high-intensity erosion distributed sporadi -
cally. In 2020, the proportion of very slight erosion area in all high-risk units declined synchronously, while the
areas of very severe and extreme erosion expanded significantly. This indicates that the inherent high sediment
yield potential of soft rock and soil units is highly susceptible to activation under extreme climate disturbances.
From the perspective of the spatial distribution of single-factor sensitivity (Fig. S3), rainfall erosivity (R) sen-
sitivity was generally high in 2010, 2020, and 2024, with the highest overall sensitivity in 2020. High-sensitivity
areas were mainly distributed in the southeastern part of the basin (Fig. S3(a)-(f)), closely related to the spatio -
temporal distribution of regional rainfall (Fig. S2). Vegetation cover ( C) sensitivity showed an overall decreas -
ing trend, but local high-sensitivity patches still existed in the loess hilly-gully region of the middle and lower
reaches (Fig. S3(g)-(l)). High-sensitivity areas of soil erodibility (K) were highly consistent with the distribution
Fig. 5 Bar chart comparing
the multi-year average soil
erosion modulus across dif-
ferent geological units.

Fig. 4 Sankey diagram of soil erosion intensity transition from 2000 to 2024.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
of argillaceous rock and aeolian soil (Fig. 2a, Fig. S3(m)). Slope (LS) sensitivity was highest in the loess hilly-
gully region, consistent with the concentrated distribution of steep slopes (Fig. 2a, c, Fig. S3(n)). The spatial
superposition of single-factor sensitivities explains the differences in erosion risk among different geological
units. For example, the aeolian loess unit exhibited both high K-factor sensitivity and high LS-factor sensitivity,
while the argillaceous rock unit showed high K-factor sensitivity and high C-factor sensitivity (poor vegetation
cover). Together, these factors account for the persistently high erosion modulus.
Quantitative contributions and interactive effects of erosion driving factors
Spearman correlation analysis
The Spearman correlation coefficients between each independent variable and the soil erosion modulus are shown
in Table S3. NDVI exhibited a highly significant negative correlation with the erosion modulus ( rs = −0.716,
p < 0.001), which was the strongest correlation among all variables. Slope showed a significant positive correla -
tion (rs = 0.328, p < 0.001). Distance to roads (rs = 0.118), land use/land cover (LULC, rs = 0.109), soil thickness
(rs = −0.078), potential evapotranspiration (PET, rs = −0.065), and lithology (rs = −0.029) were all significant but
weakly correlated. Precipitation (rs = −0.008, p = 0.158), distance to water (Dist_water, rs = 0.000, p = 0.949), and
aspect (rs = 0.008, p = 0.155) showed no significant correlation with the erosion modulus.
Highly correlated variable pairs among independent variables (| rs| > 0.3) included: Precipitation and Lithol -
ogy (rs = −0.541), PET and Precipitation ( rs = 0.558), Precipitation and Slope ( rs = 0.416), PET and Lithology
Fig. 6 Stacked bar chart of soil erosion intensity across different geological units.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
(rs = −0.362), PET and Dist_water (rs = −0.357), PET and Soil_thickness (r = 0.344), and NDVI and LULC (rs =
−0.302). No severe pairwise collinearity with |rs| > 0.7 was observed.
VIF analysis
VIF was calculated for all independent variables, and the results are shown in Table S4. The VIF values of all
variables were below 2.3, far below the severe collinearity threshold ( VIF > 10) and the moderate collinearity
threshold (VIF > 5). The variables with the highest VIF values were Precipitation (2.298) and PET (2.150), while
the VIF of NDVI was 1.165 and that of Slope was 1.533. This indicates that no multicollinearity exists among the
independent variables, and all variables can be simultaneously included in the XGBoost model.
Controlled experiments and SHAP analysis
The results of the XGBoost-SHAP model analysis indicated that the contributions of individual factors to erosion
intensity exhibited a distinct gradient (Fig. 7). In the baseline model M0, NDVI and slope ranked as the first and
second dominant factors, respectively, with SHAP value ranges far exceeding those of other variables, indicat -
ing that they possess the strongest explanatory power for the spatial heterogeneity of erosion. Specifically, high
NDVI values (red scatter points) were concentrated in the negative SHAP region, strongly inhibiting erosion;
whereas low NDVI values (blue scatter points) corresponded to high positive SHAP values, substantially intensi-
fying sediment yield. The effect direction of slope was opposite to that of NDVI, with high slope corresponding
to positive SHAP values, and steep terrain significantly amplifying erosion intensity. Factors such as land use
type, precipitation, potential evapotranspiration, distance to roads, surface substrate type, and soil thickness made
moderate contributions; the SHAP value distributions of distance to water and aspect were narrow, indicating
relatively limited independent driving effects.
The model performance and SHAP importance of the seven groups of controlled experiments are shown in
Table S6. Under random holdout validation, the full-variable model M0 achieved an R² of 0.694 and an RMSE
of 7.694 in reproducing the RUSLE estimates, with NDVI (SHAP = 5.076) and Slope (SHAP = 2.649) being the
most important predictors. After removing NDVI (M1), R2 dropped sharply to 0.127, RMSE increased to 12.677,
Fig. 7 SHAP beeswarm plot of
contribution differences among
soil erosion driving factors.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
slope became the most important variable (SHAP = 2.574), and the SHAP importance of precipitation and land
use/land cover (LULC) increased. After removing slope (M2), R2 decreased to 0.498, and NDVI remained the
most important variable (SHAP = 5.109). After removing both NDVI and slope (M3), R2 further decreased to
0.043, and the model nearly failed, with LULC and distance to water becoming the main variables. After remov-
ing all RUSLE direct factors (M4), R2 was − 0.011, lower than the mean prediction. M5, which retained only
external variables, produced exactly the same results as M4 (R2 = −0.011), indicating that LULC and precipitation
contributed zero in M4. After permuting NDVI and slope (M6), R2 was 0.029, and their SHAP importance com-
pletely disappeared. The Spearman rank correlation coefficients of SHAP importance ranking changes (relative
to M0) were: M1 (p = 0.581), M2 (p = 0.635), M3 (p = 0.560), M4 (p = 0.547), M5 (p = 0.547), and M6 (p = 0.567),
indicating that removing or permuting key variables led to significant changes in variable importance rankings.
SHAP dependence plots further reveal the nonlinear regulatory details of key factors (Fig. 8). NDVI shows
a clear threshold effect: when NDVI < 0.2, SHAP values rise sharply and erosion risk surges abruptly; when
Fig. 8 SHAP dependence plots for driving factors of soil erosion.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
NDVI > 0.2, SHAP values turn negative and tend to stabilize, with the inhibitory effect of vegetation fully acti -
vated. Slope gradient exerts mainly negative or neutral contributions in the range of 0°–10°, and its positive
contribution increases continuously and linearly beyond 10°. In terms of interactive effects, under low NDVI
conditions, red scatter points representing high slope gradients are concentrated, verifying that the combination
of “steep slope + low vegetation coverage” is the most hazardous geomorphic assemblage triggering extreme
erosion. Among land use types, cropland, forest, water bodies and impervious surfaces mostly show negative
SHAP values, while bare land presents extremely high positive SHAP values, and grassland makes a nearly
neutral contribution. For geological lithology, units of argillaceous rock, aeolian soil, and medium-fine clastic
rock interbedded with argillaceous rock show significant positive contributions, whereas lacustrine soil and intact
coarse clastic rock mostly exhibit negative SHAP values. When soil thickness exceeds 200 cm, pixels with posi-
tive contributions increase sharply, as abundant erodible source materials are provided for erosion processes. The
SHAP values of precipitation and potential evapotranspiration remain stable near zero across the entire domain,
with only a slight positive shift in the extreme high-value range, indicating that within the current range of cli -
mate fluctuations, the direct sensitivity of erosion to climatic factors is weaker than its dependence on underlying
surface factors. The NDVI threshold (0.2) and slope threshold (10°) were identified based on the inflection points
of the SHAP dependence curves; however, as they are data-driven thresholds, further validation with independent
data is needed in future studies.
Discussion
Close association between geological substrate and erosion spatial pattern
A core finding of this study is that lithology and rock-soil mass structure exert a dominant background constraint
on the regional erosion intensity pattern. The difference in multi-year average soil erosion modulus among dif -
ferent geological units can reach an order of magnitude, and this difference remained stable throughout the long
time series from 2000 to 2024, without undergoing fundamental changes due to overall vegetation improve -
ment or interannual climate fluctuations. This phenomenon indicates that geological substrate, by controlling
soil erodibility, weathering-driven sediment yield capacity, and surface sediment supply flux, is closely related
to the potential baseline of erosion intensity in each region. The aeolian loess and argillaceous rock units serve
as persistent erosion hotspots in the basin, not only due to their inherent properties of loose soil structure 54, low
cohesion55 and high susceptibility to disintegration upon wetting 56,57, but more importantly because they can
continuously supply abundant loose detritus under steep slope conditions 58, resulting in a condition that may
constrain management effectiveness for erosion reduction via vegetation restoration and engineering measures on
such substrates. Studies of soil erosion in similar semi-arid catchments at the global scale have also shown that
the inherent properties of rock and soil, such as texture, permeability, and clay mineral content, which determine
swelling-shrinkage potential, are important factors influencing soil erosion, and that soil loss estimated using
the USLE equation is far lower than that measured by the 137Cs isotope tracing method59. Studies on steep-slope
erosion in the Loess Plateau by Liu et al. (1994) also confirmed that the amplification effect of slope gradient on
erosion is far stronger in soils developed from loose parent material than in consolidated bedrock areas33, further
corroborating the synergistic association between geological substrate and topographic factors. In an erosion sen-
sitivity assessment of the coarse and sediment-rich region in the middle reaches of the Yellow River, Wang et al.
(2024) also pointed out that the soil erodibility factor is the core factor controlling regional soil erosion, followed
by topographic factors60. As an external dynamic factor, rainfall further exacerbates soil erosion risk in areas with
poor soil erosion resistance and large topographic relief, which forms cross-scale corroboration with the conclu-
sions of this study. Kou et al. (2021) confirmed through field experiments under controlled rainfall intensity that
rill development is closely related to slope microtopographic changes, and its rate of change is jointly regulated
by slope gradient and soil erosion resistance 27. In 2024, they further revealed using UA V photogrammetry that

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
terraced slope erosion is sensitive to changes in hydrological pathways, and microtopographic changes can sig -
nificantly affect erosion-deposition patterns28. This process-scale evidence provides important support for under-
standing the synergistic regulation mechanism among geological substrate, slope, and vegetation.
Mechanistically, for the widely distributed Jurassic argillaceous rock and sandstone-mudstone interbeds (Pisha
sandstone type) in the Kuye River Basin, the shear strength of undisturbed soil mass can decrease by 40%~60%
after saturation13. The surface weathered layer remains in a loose state year-round, providing sufficient sediment
sources for sheet erosion and gully erosion12. Since the initial development of the USLE model, it has been rec -
ognized that parent rock properties are the core foundation determining hierarchical differences in soil erosion
resistance32, yet their role is often masked by the dynamic changes of surface factors such as vegetation and
topography61. By incorporating surface substrate type into the driving factor system, this study shows a persistent
spatial association with estimated water erosion, which compensates for the deficiency of conventional driving
factor analyses that only consider external influencing factors such as surface soil, topography and rainfall in char-
acterizing geological background constraints. This also implies that geological substrate should not be regarded
as an ordinary input variable in erosion models, but rather understood as the fundamental framework of erosion
spatial differentiation. This understanding is consistent with the “theory of geological background constraint”
proposed by Yang et al. (2024) in a study on erosion vulnerability in southwestern Hubei, jointly highlighting the
important background association of geological conditions in different erosion type zones35. Future studies could
draw on the multi-scenario simulation method adopted by Wu et al. (2024a) in ecological risk assessment of the
Kuye River Basin20, as well as the morphology-stratified assessment framework proposed by Yuan et al. (2026),
to formulate differentiated governance strategies for different geological units52.
Synergistic effects of multiple factors and the triggering effect of extreme climate
The driving mechanism analysis based on the XGBoost-SHAP model demonstrates that individual driving fac -
tors do not act independently, but form a hierarchical regulatory chain centered on “geological substrate – slope
gradient – vegetation”. Among them, the geological substrate provides the sediment source basis and determines
the upper limit of erosion potential; slope gradient supplies hydrodynamic conditions and governs the spatial dis-
tribution pattern of erosion; vegetation exerts a regulatory role through raindrop interception, runoff retardation
and soil consolidation, serving as the dominant factor modulating the dynamic changes of erosion. The coupling
relationship among the three collectively determines the basic erosion pattern of the basin, while human activities
(e.g., land use change, road construction, and mineral exploitation) exert superimposed impacts at the local scale,
further regulating or disturbing this pattern by altering surface cover and rock-soil structure 62. This is consistent
with conclusions from multi-regional erosion driving studies worldwide: natural factors underpin the spatial pat-
tern of erosion 63, whereas anthropogenic activities modulate the intensity and dynamics of erosion by altering
underlying surface conditions2.
A study based on the Geodetector model by Liu et al. (2025) confirmed that the interaction between fractional
vegetation coverage and slope gradient exhibits a significant “two-factor enhancement effect”, with its explanatory
power for the spatial differentiation of erosion far exceeding that of either single factor36. This study preliminarily
indicates that the strength of this interaction may be influenced by the geological substrate, but further verification
using stratified models or interaction SHAP values is needed. Regarding the impacts of human activities, Liu et
al. (2023) pointed out that activities such as coal mining and road construction in the Kuye River Basin aggravate
local erosion primarily by destroying the surface rock-soil structure and reducing vegetation coverage 15, acting
as superimposed disturbances on the natural erosion pattern. This is consistent with the conclusion of this study
that human activity factors rank moderately in contribution and mainly exert local modulating effects.
In combination with the land use type map, cultivated land and built-up land are concentrated in the loess
hilly-gully region of the middle and lower reaches, where vegetation cover is relatively low, consistent with the
high-sensitivity areas of the C factor (Fig. 2d; Fig. S3(g)–(l)). During the extreme rainfall in 2020, the sensitiv -
ity of the R factor increased sharply and overlapped with the high-sensitivity areas of the C factor, leading to a

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
synchronous rebound in erosion intensity of soft substrate units in the middle and lower reaches. This indicates
that under extreme climatic conditions, the combination of enhanced rainfall erosivity, insufficient vegetation
cover, and fragile geological substrate is the key driver of intensified erosion. Kou et al. (2024b), using SBAS-
InSAR technology, monitored significant erosion-deposition dynamic adjustments in rill areas of loess tablelands
after extreme rainfall, confirming the strong triggering effect of extreme climatic events on soft geological units29;
Kou et al. (2026) further showed in a study of the Three Gorges Reservoir area that the accelerated response of
landslide deformation lagged behind extreme rainfall by 3 ~ 5 days, reflecting the lag effect of pore-water pres -
sure diffusion53. This suggests that the lag effect of moisture transmission should be considered in the analysis of
erosion driving mechanisms.
Under normal climatic conditions, existing soil and water conservation measures can effectively curb the sedi-
ment yield rate of soft rock-soil units and maintain the overall downward trend of erosion intensity. However,
when high-intensity extreme rainfall events occur, rainfall erosivity far exceeds the protection threshold of veg -
etation and engineering measures within a short period. Climatic factors temporarily rise to become the dominant
controlling factors, directly penetrating the buffer barrier of the ecosystem and rapidly activating high-intensity
erosion in areas with fragile rock-soil substrates 64. A global-scale review by Li and Fang (2016) also noted that
the stimulating effect of extreme rainfall events on soil erosion exhibits prominent substrate dependence 65, and
a soft and loose material base is a prerequisite for severe erosion triggered by extreme events. Long-term obser -
vations in the Wangmaogou small watershed of the Loess Plateau by Shi et al. (2024) also found that the basin
erosion modulus in extreme rainfall years can reach 2 ~ 3 times that in normal years 5, and the newly increased
high-intensity erosion is almost entirely concentrated on loess and mudstone slopes with poor erosion resistance,
which is highly consistent with the spatial pattern identified in this study. Field investigations after the “7·26”
extreme rainstorm in the Wuding River in 2017 further showed that the gully erosion expansion rate in sand-
mudstone outcrop areas can reach 5 ~ 8 times the annual average, and some silted-up check dams experienced
dam failure and re-erosion 66, confirming the strong triggering effect of extreme rainfall on areas with fragile
geological substrates.
This phenomenon reveals a stark reality: current ecological restoration measures have not yet fundamentally
altered the inherent sediment yield potential of the geological substrate. Simulation research by Eekhout and
de Vente (2019) indicated that the increasing frequency of extreme rainfall under future climate scenarios will
cause a far greater erosion rebound magnitude in high geological risk areas than in low-risk areas, highlighting
the importance of erosion risk assessments that incorporate geological background considerations67. Against the
backdrop of global climate change, the frequency of extreme rainstorms is on the rise, and regional soil and water
loss still faces a systematic rebound risk68.
Circular reasoning risk in coupled XGBoost-SHAP and RUSLE studies
The controlled experiments directly confirmed the structural overlap between RUSLE internal factors and
XGBoost independent variables. NDVI is a direct proxy variable for the C factor (NDVI→FVC→C→A), and
slope is a direct proxy variable for the LS factor (Slope→S→A). In the baseline model M0, the SHAP importance
of NDVI and slope was far higher than that of other variables (5.076 and 2.649, respectively), but this mainly
arises from the mathematical construction relationship of the RUSLE equation. After removing NDVI, the model
R2 plummeted from 0.694 to 0.127 (an 82% decrease), indicating that the high performance of XGBoost almost
entirely depends on NDVI. After removing both NDVI and slope, R2 dropped to 0.043, and the model nearly
failed. When all RUSLE direct factors ( R, C, LS, and P) were removed, R2 was − 0.011, lower than the mean
prediction. After permuting NDVI and slope, their SHAP importance completely disappeared. These results con-
sistently indicate that the XGBoost model mainly reproduces the spatial structure of the RUSLE equation rather
than independently predicting actual erosion processes.
Based on the above results, this study interprets SHAP values as the “covariate importance of the RUSLE-esti-
mated erosion pattern” rather than the “causal driving mechanisms of actual erosion processes.” The high SHAP

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
importance of NDVI and slope should not be interpreted as “vegetation and topography are the dominant driving
factors of erosion,” but rather as “within the RUSLE framework, the C factor and LS factor contribute most to the
spatial pattern of the erosion modulus.” This distinction is crucial for avoiding circular reasoning. SHAP values
describe model associations, not causal effects.
All non-RUSLE variables (PET, Dist_road, Dist_water, Aaspect, Soil_thickness, Lithology) had no indepen -
dent explanatory power for the spatial pattern of the RUSLE-estimated erosion modulus (M5 R2 = −0.011).
This may be because the RUSLE model itself does not include these processes (e.g., wind erosion, gravita -
tional erosion, road erosion), or because the spatial variation of these variables is insufficient to explain the pat -
tern of RUSLE outputs. Although VIF analysis indicated no multicollinearity among independent variables (all
VIF < 2.3), VIF only detects collinearity among independent variables and cannot detect the structural relation -
ship between independent variables and the dependent variable. Therefore, VIF results cannot replace controlled
experiments.
Implications for precision management and uncertainty analysis
The driving mechanism analysis results of this study provide clear management implications for soil and water
conservation in the Kuye River Basin. Given the persistent spatial association of geological substrate with ero -
sion intensity, future soil and water conservation efforts should shift from the current homogeneous governance
model based on administrative units and geomorphic types to a “geological-unit-based zoned precision imple -
mentation strategy”, so as to improve the cost-effectiveness of governance investment.
Specifically, on the basis of systematically identifying the fine spatial distribution, physical and chemical prop-
erties of rock and soil, weathering rates and sediment transport rules of high-risk geological units such as aeolian
soil, argillaceous rock, and clastic rock interbedded with argillaceous rock, differentiated governance strategies
should be formulated for geological units with different risk levels: ① For high-risk soft substrate units, engineer-
ing measures should be taken as the priority to construct a multi-level protection system of “slope soil stabili -
zation – gully sediment retention – gully outlet flood control”, combined with barren-tolerant native shrub and
herb species with developed root systems to rapidly improve soil shear strength, with a focus on curbing water
erosion and gully expansion11; ② For medium-risk geological units, natural ecological restoration should be the
main approach, supplemented by local engineering measures such as slope terracing and gully check dams, and
the soil and water conservation function should be enhanced through natural vegetation succession; ③ For low-
risk geological units, prevention and protection should be the priority, with strict control over development and
construction activities to maintain the existing favorable ecological conditions. This differentiated governance
approach is also consistent with the direction of “classified and zoned precise management and control” for the
Kuye River Basin proposed by Chen et al.69.
Meanwhile, the uncertainties and limitations of this study should also be recognized. First, the surface sub -
strate classification in this study is based on a 1:200,000 regional geological map, which has limited spatial accu-
racy. It is difficult to characterize meso-structural differences such as rock weathering degree and joint fracture
development at the regional scale, which may lead to certain deviations in the identification of local high erosion
risk areas. This is also a common limitation of empirical models in regional erosion assessment: homogenization
treatment at the grid scale cannot fully reflect the spatial heterogeneity of rock and soil properties34. A compara-
tive study by Nearing et al. (2000) also pointed out that both the RUSLE and WEPP models experience vary -
ing degrees of decline in simulation accuracy in transitional zone watersheds with complex lithology 70, which
requires further calibration with in-situ monitoring data in the future. Second, the RUSLE model adopted in this
study takes water erosion as the core simulation object, and does not systematically incorporate the contributions
of wind erosion, freeze-thaw erosion and gravitational erosion. The northern part of the Kuye River Basin is
located in the transition zone between the Mu Us Sandy Land and the Loess Plateau, where the alternating action
of wind and water erosion is prominent, which may underestimate the total erosion intensity of this region to a
certain extent26. Third, the characterization of human activities in the driving factor system is still relatively rough.

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
Refined indicators such as coal mining intensity, number and siltation status of check dams, and investment in soil
and water conservation projects have not been included in the analysis, making it difficult to fully quantify the
bidirectional effects of anthropogenic governance and anthropogenic disturbance. Fourth, the XGBoost model
predicts RUSLE-estimated outputs, and SHAP values describe model associations rather than causal effects.
Moreover, controlled experiments confirmed that the high importance of NDVI and slope partly arises from the
structure of the RUSLE equation. Additionally, the random 80%/20% splitting and five-fold cross-validation used
in this study primarily assess the model’s capacity to reconstruct RUSLE outputs within the study domain. Due to
spatial autocorrelation, spatially adjacent samples may lead to potentially optimistic performance estimates, and
the reported model performance (R² = 0.694) does not directly establish predictive transferability across different
regions or years. Future studies need to integrate independent observational data (e.g., check dam sedimentation,
137Cs tracing, and sediment discharge at hydrological stations) to establish predictive models, so as to distinguish
RUSLE structural associations from actual erosion drivers. Follow-up studies can combine high-resolution geo-
logical survey data, multi-source in-situ monitoring data and multi-process coupled erosion models to further
improve the accuracy of erosion mechanism analysis and the pertinence of governance recommendations.
Conclusion
Overall, this study systematically examined the spatiotemporal evolution of soil erosion and its associated driving
factors in the Kuye River Basin from 2000 to 2024 by integrating the RUSLE model with an XGBoost-SHAP
coupled analytical framework. The results indicate that although the overall soil erosion intensity in the basin
showed an improving trend, its spatial pattern remained stable, exhibiting a pronounced gradient differentiation
associated with geological lithology. Aeolian loess, argillaceous rock, and clastic rock interbedded with argil -
laceous rock units were identified as persistent high erosion risk zones, with their multi-year average soil ero -
sion modulus and areal proportion of extreme erosion significantly higher than those of other units. Geological
substrate was the background factor most closely related to the spatial differentiation of regional water erosion
intensity in this dataset: the spatial distribution of single-factor sensitivity showed that high-sensitivity areas of
the K factor coincided with the distribution of argillaceous rock and aeolian sand, high-sensitivity areas of the LS
factor were consistent with steep-slope concentration zones in the loess hilly-gully region, and high-sensitivity
areas of the C factor overlapped with low-vegetation-cover areas in the middle and lower reaches. The spatial
superposition of these three factors is closely associated with the persistently high erosion risk in aeolian loess
and argillaceous rock units. Moreover, soft rock-soil units not only possess high inherent erosion potential but
also exhibit extreme sensitivity to extreme rainfall events; the basin-wide erosion rebound in 2020 highlights
that current vegetation restoration measures may have limited effectiveness in coping with extreme climate per -
turbations. Within the RUSLE framework, NDVI and Slope, corresponding to the C and LS factors, contributed
most to the spatial pattern of erosion modulus (with SHAP value ranges far exceeding those of other variables).
However, controlled experiments indicated that this result partly arises from the mathematical structural relation-
ship of the RUSLE equation, and the model performance (R2 = 0.694) primarily reflects its capacity to reconstruct
RUSLE outputs within the study domain rather than cross-regional predictive transferability. External environ -
mental variables (potential evapotranspiration, distance to roads, distance to water systems, aspect, soil thickness,
and lithology) showed limited independent explanatory power for RUSLE outputs (M5 R2 = −0.011). Therefore,
the SHAP results of this study should be interpreted as the “covariate importance of the RUSLE-estimated ero -
sion pattern” rather than the “causal mechanisms of actual erosion processes.” Accordingly, it is recommended
that geological unit differentiation be incorporated as a core consideration in future soil and water conservation
planning, and that a precision governance system based on zoning of rock-soil erosion resistance be established to
enhance the scientific rationality and long-term effectiveness of erosion control measures. Future studies should
integrate independent observational data (e.g., check dam sedimentation, 137Cs tracing, and sediment discharge

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
at hydrological stations) to establish predictive models for differentiated geological units, so as to distinguish
RUSLE structural associations from actual erosion drivers.
Supplementary Information The online version contains supplementary material available at https://doi.org/10.1038/
s41598-026-79921-0.
Acknowledgements This work was supported by the China Geological Survey (No. DD202607102104, DD202605101904,
DD20242461) and the National Natural Science Foundation of China (No. U2344227).
Author contributions XBL. and YLZ.wrote the main manuscript text and prepared the figures; CXK.and JSX. and HZ and
CHH. designed the research plan, data collection, and performed data analysis; DXX. contributed to the conceptualization
of the manuscript; ZQY . reviewed the manuscript; JHZ. contributed to the conceptualization, provided funding support, and
reviewed the manuscript. All authors reviewed and revised the manuscript.
Data availability The meteorological data used in this study were derived from the Basic Meteorological Element Obser -
vation Dataset of China’s National Surface Meteorological Stations ( h t t p s : / / d a t a . c m a . c n / d a t a / c d c d e t a i l / d a t a C o d e / A . 0
0 1 2 . 0 0 0 1 . h t m l ) . The land use data were obtained from the China Land Cover Dataset (CLCD, https://doi.org/10.5281/
zenodo.18180184). The potential evapotranspiration data came from the 1 km Monthly Potential Evapotranspiration Dataset
of China ( h t t p s : / / w w w . g e o d a t a . c n / d a t a / d a t a d e t a i l s . h t m l ? dataguid=34595274939620&docId=2152). The Normalized Dif -
ference Vegetation Index (NDVI) data were sourced from the 30 m NDVI Dataset of China’s Loess Plateau ( http://loess.
geodata.cn). The DEM data were extracted from the 30 m Aster GDEM Digital Elevation Model hosted on the Geospatial
Data Cloud ( h t t p s : / / w w w . g s c l o u d . c n / ) . The soil property data were acquired from the High-Resolution National Soil Grid
Basic Attribute Dataset of China (https://www.geodata.cn/data/datadetails.html?dataguid=36810085119113&docId=0). The
stratigraphic lithology data were retrieved from the spatial database of the public-version 1:200,000 national digital geo -
logical map ( h t t p : / / d c c . c g s . g o v . c n ) . All intermediate datasets and results generated during this study are available from the
corresponding author upon reasonable request.
Declarations
Competing interests The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 Interna-
tional License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as
long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence,
and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material
derived from this article or parts of it. The images or other third party material in this article are included in the article’s Cre-
ative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s
Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you
will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit h t t p : / / c r e a t i v e c o m m o
n s . o r g / l i c e n s e s / b y - n c - n d / 4 . 0 / .
References

  1. Kumar, M. et al. A global review of rainfall erosivity estimation: methods, challenges, and way forward. J. Earth Syst.
    Environ. 10, 8149–8180 (2026).
  2. Bayouli, O. T., Zhang, W. & Bayouli, H. T. Combining RUSLE model and the vegetation health index to unravel the
    relationship between soil erosion and droughts in southeastern Tunisia. J. J. Arid Land. 15 (11), 1269–1289 (2023).
  3. Shangguan, Z. P. Achievements and challenges of comprehensive control of soil and water loss on the Loess Plateau
    under the new situation. J. J. Shandong Univ. (Nat. Sci.) 61(9), 1–7. https://doi.org/10.6040/j.issn.1671-9352.0.2026.152
    (2026).
  4. Zhang, Z. P. et al. Analysis of spatial and temporal evolution and dynamic driving force of soil water erosion inthe middle
    reaches of the Yellow river in the rich and coarse sediment area. J. J. Soil. Water Conserv. 38 (2), 85–96. h t t p s : / / d o i . o r g /
    1 0 . 1 3 8 7 0 / j . c n k i . s t b c x b . 2 0 2 4 . 0 2 . 0 1 4 (2024).
  5. Shi, X. J. et al. Spatial-temporal evolution of soil erosion in a typical watershed of the Loess Plateau. J. Acta Geogr.
    Sinica. 79 (7), 1787–1803. https://doi.org/10.11821/dlxb202407010 (2024).

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
6. Deng, X. Z., Liu, Y . J. & Liu, Y . H. Climate-vegetation threshold mechanism and temporal response characteristics of soil
erosion in the Yellow River Basin. J. Yellow River. 47 (9), 77–84. https://doi.org/10.3969/j.issn.1000-1379.2025.09.011
(2025).
7. Tang, Q. H. et al. Development of an eco-hydrological model for flood and drought risk assessment under a changing
environmentin the middle reaches of the Yellow River. J. Acta Geogr. Sinica. 78 (7), 1666–1676. h t t p s : / / d o i . o r g / 1 0 . 1 1 8 2
1 / d l x b 2 0 2 3 0 7 0 0 8 (2023).
8. Fu, B. J. et al. Current conditions, issues, and suggestions for ecological protection and high-quality development in
Loess Plateau. J. Bull. Chin. Acad. Sci. 38 (8), 1110–1117. https://doi.org/10.16418/j.issn.1000-3045.20230525001
(2023).
9. Deng, L. & Shangguan, Z. P. High quality developmental approach for soil and water conservation and ecological
protection on the Loess Plateau. Front. Agric. Sci. Eng. 8(4), 501. (2021).
10. Zhang, P. et al. Researeh on curent situation, problems, and countereasures of Soil erosion models on Loess Plateau. J.
Res. Soil. Water Conserv. 33(1), 1–10. https://doi.org/10.13869/j.cnki.rswc.2026.01.042 (2026).
11. Zhang, P. & Yao, W. Y . Equilibrium theory and key technologies for optimal layout of Yudiba dam systems in the new
era. J. Water Resour. (7), 24–31. https://doi.org/10.3969/j.issn.1000-1123.2026.07.003 (2026).
12. Gao, H. Y . et al. Research progress on soil erosion and ecological restoration in the Kuye River Basin of the middle
Yellow River. J. J. Inner Mongolia Univ. (Nat.l Sci. Ed.) 55(3), 329–336. https://doi.org/10.13484/j.nmgdxxbzk.20240312
(2024).
13. Qin, F. C., Yang, Z. Q. & Li, L. Research progress on soil erosion mechanism and ecological restoration technology in
feldspathic sandstone region. J. J. Beijing For. Univ. 42(12), 142–150. https://doi.org/10.12171/j.1000-1522.20190430
(2020).
14. Zhang, D. et al. Impacts of coal mining activities on dissolved sulfate in the Kuye River Basin, the Midstream of Yellow
River, China. J. J. Earth Sci. Environ. 45(2), 414–426. https://doi.org/10.19814/j.jese.2022.10029 (2023).
15. Liu, Q. et al. Landscape pattern evolution and driving forces of land use in Kuye River Basin from 1980 to 2020. J. Res.
Soil. Water Conserv. 30(5), 335–341. https://doi.org/10.13869/j.cnki.rswc.2023.05.029 (2023).
16. He, Y . et al. Runoff variation and influencing factors in the Kuye River Basin of the middle Yellow River. Front. Environ.
Sci. 10, 877535. https://doi.org/10.3389/fenvs.2022.877535 (2022).
17. Li, S. et al. Underground mining impact on groundwater in Kuye River Basin, China: A coupling model study. River 4,
106–115. https://doi.org/10.1002/rvr2.70000 (2025).
18. Sun, G. X. et al. Future runoff responses to dynamic coal mining expansion and climate change in China’s Kuye river basin
using a modified hydrological model. J. J. Hydrology: Reg. Stud. 67, 103873. https://doi.org/10.1016/j.ejrh.2026.103873
(2026).
19. Guo, X. Y . et al. Impacts of climate change and human activities on vegetation cover during the growing season in the
Kuye River Basin. J. Trans. Chin. Soc. Agric.l Eng. (Trans. CSAE) 41(10), 304–313. h t t p s : / / d o i . o r g / 1 0 . 1 1 9 7 5 / j . i s s n . 1 0 0
2 - 6 8 1 9 . 2 0 2 4 0 5 1 8 4 (2025) ((in Chinese with English abstract)).
20. Wu, Y . H. et al. Modelling ecological hazards and causal factors in the Yellow River Basin’s key tributaries: A case study
of the Kuye River Basin and its future outlook. Sustainability 16(16), 6977. https://doi.org/10.3390/su16166977 (2024a).
21. Wu, Y . H. et al. Exploring landscape ecological risk with human activity intensity and correlation in the Kuye River
Basin. J. Front. Ecol. Evol. 12, 1409515. https://doi.org/10.3389/fevo.2024.1409515 (2024b).
22. Wu, Y . H. et al. Investigating the spatio-temporal evolution of land cover and ecosystem service value in the Kuye River
Basin. Water 16(17), 2456. https://doi.org/10.3390/w16172456 (2024c).
23. Wu, Y . H. et al. Construction and optimisation of watershed scale ecological network: a case study of Kuye River Basin.
J. Front. Environ. Sci. 12, 1364568. https://doi.org/10.3389/fenvs.2024.1364568 (2024d).
24. Liu, Q., Yu, F. H. & Mu, X. M. Evaluation of the ecological environment quality of the Kuye River Source Basin using
the remote sensing ecological index. Int. J. Environ. Res. Public Health 19(19), 12500. h t t p s : / / d o i . o r g / 1 0 . 3 3 9 0 / i j e r p h 1 9
1 9 1 2 5 0 0 (2022).
25. Chen, L. L. The analysis of LUCC, erosion responses and human activites contribution in the Kuye River Basin (Northwest
A&F University, 2015).
26. He, Y . et al. Spatio-temporal pattern and trend prediction of soil erosion in Kuye River Basin. J. J. Soil. Water Conserv.
38 (5), 10–19. https://doi.org/10.13870/j.cnki.stbcxb.2024.05.028 (2024).
27. Kou, P. L. et al. Rill development and its change rate: A field experiment under constant rainfall intensity. Catena 199,
105112. https://doi.org/10.1016/j.catena.2020.105112 (2021).
28. Kou, P. L. et al. Assessment of loess terraced slope erosion using unmanned aerial vehicle photogrammetry: Topographic
and hydrological effects. Land Degrad. Dev. 35(7), 2462–2475. https://doi.org/10.1002/ldr.5073 (2024a).
29. Kou, P. L. et al. Analyzing gully erosion and deposition patterns in loess tableland: Insights from small baseline subset
interferometric synthetic aperture radar (SBAS InSAR). J. Sci. Total Environ. 916, 169873. h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . s c i
t o t e n v . 2 0 2 4 . 1 6 9 8 7 3 (2024b).
30. Li, Z. B., Zhu, B. B. & Li, P. Advancement in study on soil erosion and soil and water conservation. J. Acta Pedol. Sinica.
45 (5), 802–809 (2008).

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
31. Zheng, F. L. et al. Review of research progress in soil erosion prediction model. J. Bull. Soil. Water Conserv. 21 (6),
16–1832 (2001).
32. Wischmeier, W. H. & Smith, D. D. Predicting rainfall erosion losses: a guide to conservation planning (U.S. Department
of Agriculture, Science and Education Administration, 1978).
33. Liu, B. Y ., Nearing, M. A. & Risse, L. M. Slope gradient effects on soil loss for steep slopes. J. Trans. ASAE. 37 (6),
1835–1840 (1994).
34. Renard, K. G. Predicting soil erosion by water: a guide to conservation planning with the Revised Universal Soil Loss
Equation (RUSLE) (U.S. Department of Agriculture, Agricultural Research Service, 1997).
35. Yang, Y . P. et al. Vulnerability assessment of soil eroson in sou hwestern Hubei Province based on RUSLE model
and geographic detector. J. J. Soil. Water Conserv. 38 (1), 91–103. https://doi.org/10.13870/j.cnki.stbcxb.2024.01.015
(2024).
36. Liu, Y . H. et al. Changes of soil erosion based on RUSLE model and driving factors in Shanxi section of Yellow River
Basin. J. Soils. 57 (2), 452–460. https://doi.org/10.13758/j.cnki.tr.2025.02.024 (2025).
37. Zhang, X. et al. Spatiotemporal dynamics of vegetation in the Beijing-Tianjin-Hebei region during 1986–2024: a driving
mechanism analysis based on XGBoost-SHAP. J. Clim. Change Res. 22 (3), 316–330. h t t p s : / / d o i . o r g / 1 0 . 1 2 0 0 6 / j . i s s n . 1 6
7 3 - 1 7 1 9 . 2 0 2 6 . 0 2 9 (2026).
38. Tarek, Z. et al. Soil erosion status prediction using a novel random forest model optimized by random search method. J.
Sustain. 15(9), 1–18 (2023).
39. Bai, Y . et al. Identifying spatiotemporal dynamics and driving factors of landscape pattern in Giant Panda National Park
(GPNP) using an interpretable machine learning technology. J. Ecol. Indic. 178, 114121 (2025).
40. Qiu, D. E. et al. Multi-scenario simulation and future planning optimization of ecological security pattern of Chengdu-
Chongqing urban agglomeration based on XGBoost-PLUS model. J. Acta Ecol. Sinica. 45 (16), 7905–7920. h t t p s : / / d o i .
o r g / 1 0 . 2 0 1 0 3 / j . s t x b . 2 0 2 5 0 1 0 9 0 0 7 9 (2025).
41. Yang, J. & Huang, X. The 30 m annual land cover datasets and its dynamics in China from 1985 to 2025. Data set. Earth
Syst. Sci. Data. 13 (1), 3907–3925 (2026).
42. Peng, S. Z. et al. 1 km monthly temperature and precipitation dataset for China from 1901 to 2017. Earth Syst. Sci. Data
11, 1931–1946. https://doi.org/10.5194/essd-11-1931-2019 (1901).
43. Cai, Y . et al. China Earth observation data cube: The 30-m seamless annual leaf-on landsat composites from 1985 to
2023. J. J. Remote Sens. https://doi.org/10.34133/remotesensing.0698 (2025).
44. Liu, F. et al. Mapping high resolution national soil information grids of China. Sci. Bull. h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / j . s c i b . 2
0 2 1 . 1 0 . 0 1 3 (2021).
45. Li, C. Y . et al. Spatial database of national 1:200,000 digital geological map (public version). J. Geol. China. 46 (S1),
1–10. https://doi.org/10.12029/gc2019Z101 (2019).
46. [34] Ministry of Water Resources of the People’s Republic of China. Classification and grading standard for soil erosion:
SL 190–2007. Standard (China Water & Power, 2008).
47. Sharpley, A. N. & Williams, J. R. EPIC-erosion/productivity impact calculator: 1. Model documentation. II: User manual
(U.S. Department of Agriculture, 1990).
48. Zhang, K. L. et al. Erodibilityof agricultural soils on the Loess Plateau of China. Soil Tillage Res. 76(2), 157–165. h t t p s
: / / d o i . o r g / 1 0 . 1 0 1 6 / j . s t i l l . 2 0 0 3 . 0 9 . 0 0 7 (2004).
49. Liu, B. Y . et al. Slope length effects on soil loss for steep slopes. Soil Sci. Soc. Am. J. 64(5), 1759–1763 (2000).
50. Cai, C. F. et al. Study of applying USLE and geographical information system IDRISI to predict soil erosion in small
watershed J. J. Soil Water Conserv. https://doi.org/10.13870/j.cnki.stbcxb.2000.02.005 (2000).
51. Zhang, L. et al. Spatiotemporal dynamics and non-linear drivers of carbon storage in the Pisha sandstone area: A coupled
PLUS-InVEST and XGBoost-SHAP framework. Sustainability 18(13), 6595. https://doi.org/10.3390/SU18136595
(2026).
52. Yuan, Z. W. et al. Enhanced landslide detection via multi-module deep learning: A morphology-dependent performance
analysis approach. Bull. Eng. Geol. Environ. https://doi.org/10.1007/s10064-026-05195-5 (2026).
53. Kou, P. L. et al. Hydro-gravitational dominance and differential landslide deformation unveiled by SBAS-InSAR-
enhanced susceptibility mapping in Wushan, Three Gorges. Nat. Hazards 122, 95. h t t p s : / / d o i . o r g / 1 0 . 1 0 0 7 / s 1 1 0 6 9 - 0 2 5 - 0
7 8 4 1 - 9 (2026).
54. Jiang, C. et al. Research status of loess hydraulic erosion-landslide-mudflow disaster chain. J. J. Nat. Disasters. 28 (1),
38–43. https://doi.org/10.13577/j.jnd.2019.0106 (2019).
55. Zhang, L. Experimental study on the slake durability of mudstone under the action of rainfall and drying. J. Sci. Technol.
Eng. 20 (15), 6221–6227 (2020).
56. Song, S. J. et al. Effects of overlying rock stratigraphy on the morphology of concave loess slopes and the effect of soil
erosion in the coal mining subsidence area of Northern Shaanxi, China. J. Metal Mine h t t p s : / / d o i . o r g / 1 0 . 1 9 6 1 4 / j . c n k i . j s
k s . 2 0 2 5 0 5 0 1 6 (2025).
57. Quan, Y . et al. From mineral dissolution to slope failure: wet-dry cycling driven multiscale degradation of Cenozoic red-
bed mudstone, NE Tibetan Plateau. J. Eng. Geol. 362, 108530 (2025).

Article in Press
Scientific Reports
https://doi.org/10.1038/s41598-026-79921-0
58. Zhu, M. Y . et al. Effects of different vegetation types and restoration processes on runoff and sediment on steep slopes of
the Loess Plateau. J. Res. Soil. Water Conserv. 30 (6), 57–66. https://doi.org/10.13869/j.cnki.rswc.2023.06.002 (2023).
59. Azaiez, N. Combined use of caesium-137 isotope and the universal soil loss equation to establish a soil loss budget along
hill slopes in El Maasra basin. J. Soil. Environ. 43 (1), 127–147. https://doi.org/10.25252/SE/2024/243161 (2024).
60. Wang, M. et al. Variation of soil erosion sensitivity in plentiful and coarse sediment area of the middle Yellow River. J.
Sci. Soil. Water Conserv. 23 (2), 55–65. https://doi.org/10.16843/j.sswc.2024061 (2025).
61. Williams, J. R., Dyke, P. T., Jones, C. A. & EPIC A new method for assessing erosion’s effect on soil productivity. J. J.
Soil. Water Conserv. 38 (5), 381–383 (1983).
62. Kuntoro, S. W. et al. Erosion extremes under changing climate: Integrating extreme rainfall erosivity in a tropical mining
landscape. J. Environ. Challenges. 24, 101574. https://doi.org/10.1016/J.ENVC.2026.101574 (2026).
63. Fang, Q. et al. Effects of simulated successive extreme rainfall events on surface and belowground runoff and soil loss
on flumes that mimic soil-mantled karst slopes. J. Catena. 272, 110311. https://doi.org/10.1016/J.CATENA.2026.110311
(2026).
64. Yan, Y . C. et al. Spatiotemporal analysis of rainfall erosivity in the Beijing-Tianjin-Hebei region. J. South-to-North Water
Transfers and Water Science & Technology, 1–17. https://link.cnki.net/urlid/13.1430.TV .20260430.1925.002 (2026).
65. Li, Z. & Fang, H. Impacts of climate change on water erosion: A review. Earth Sci. Rev. 163, 94–117 (2016).
66. Sun, C. J. et al. Analysis of soil and water loss characteristics in typical small watersheds of the middle Yellow River based
on RUSLE model. J. Southwest. China J. Agric. Sci. 35(1), 200–208. https://doi.org/10.16213/j.cnki.scjas.2022.1.026
(2022).
67. Eekhout, J. P. C. & de Vente, J. The implications of bias correction methods and climate model ensembles on soil erosion
projections under climate change. Earth Surf. Process. Landf. 44(5), 1137–1147 (2019).
68. Yin, C. et al. Future soil erosion risk in China: Differences in erosion driven by general and extreme precipitation under
climate change. J. Earth’ s Future. 13 (3), e2024EF005390 (2025).
69. Chen, Y . C. et al. Analysis and prediction of soil erosion under different land use change scenarios in the Kuye River
Basin based on PLUS CSLE model. J. Environ. Sci. 46 (8), 4971–4981. https://doi.org/10.13227/j.hjkx.202406130
(2025).
70. Nearing, M. A. et al. Modeling soil erosion and sediment yield with RUSLE, WEPP, and spatial data bases. J. J. Soil.
Water Conserv. 55 (3), 300–306 (2000).
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
Authors and Affiliations
Xinbin Li1,2,3 · Yilong Zhu1,3 · Chengxin Kang1,3 · Junsheng Xi1,3 · Hao Zhao1,3 ·
Duoxun Xu1,3 · Zhiqiang Yin4 · Chaohui Han1,3 · Jianghua Zhang1,3
Yilong Zhu
zhuyilong@mail.cgs.gov.cn
Jianghua Zhang
53235266@qq.com
1 Xi’an Mineral Resources Survey, Geological Survery, No.7 Fengqi West Road, Xi’an 710100, China
2 State Key Laboratory of Deep Earth Exploration and Imaging, School of Geophysics and Information Technology,
China University of Geosciences, Beijing 100083, China
3 Qinling-Loess Plateau Transition Zone Observation and Research Station for Coupling of Soil and Water Elements
and Conservation of Biological Resources, Tongguan 714300, Shaanxi, China
4 Natural Resources Survey, China Geological Survey, Beijing 100055, China

论文原文与出版信息

Scientific Reports
查看论文原文

继续探索