Spatial and temporal evolution and interaction of soil erosion intensity and influencing factors in Wenzhou City from 2000 to 2023

Abstract

Soil erosion constitutes a critical global environmental issue, which is particularly pressing in rapidly urbanizing coastal hilly regions where natural and anthropogenic pressures are intensely intertwined. This study takes Wenzhou City, a typical coastal area in southeastern China, as the research object. Based on multi-source remote sensing and meteorological data from 2000 to 2023, we comprehensively applied the InVEST model to assess soil erosion intensity. This was combined with hotspot analysis, Pearson correlation, and the interpretable machine learning method (XGBoost-SHAP) to quantify its spatio-temporal evolution patterns and decipher complex factor interactions. The main conclusions are as follows: (1) The erosion intensity showed phased temporal changes, transitioning from slight fluctuations before 2010 to continuous mitigation thereafter. The combined area of slight and light erosion consistently accounted for over 99% of the total, while the high-intensity erosion area drastically decreased from 27.91 km² to 2.61 km². (2) Spatially, a stable pattern of “slight erosion in the west, light erosion in the east” was observed, with significant clustering characteristics. Hotspots were concentrated in the hilly areas of the northeast, central, and southern parts, whereas cold spots were stably distributed in the western plains. (3) The interaction between slope gradient and precipitation was the core natural driving force, with correlation coefficients of 0.52–0.54 and 0.17–0.25, respectively. In contrast, soil moisture exhibited a significant and increasingly strong negative regulatory effect, with its importance surpassing that of land use type after 2020. The synergistic variation between evapotranspiration and temperature indirectly regulated soil erosion intensity.

Introduction

Soil erosion represents a prominent global ecological and environmental challenge. It undermines soil resource sustainability, reduces land productivity, degrades ecosystem functions, and constrains regional sustainable development^1,2. As a complex spatiotemporal process resulting from the intertwined effects of natural factors and human activities, soil erosion has attracted increasing academic attention under the dual pressures of rapid urbanization and global climate change^3,4. Therefore, investigating the spatiotemporal evolution patterns of soil erosion and revealing its underlying driving mechanisms has become an important research theme in interdisciplinary fields such as geography, environmental science, and sustainability studies.

Wenzhou City in Zhejiang Province, located on China’s southeastern coast, serves as a typical region for exploring soil erosion dynamics in economically active coastal hilly areas. Topographically, Wenzhou is predominantly characterized by mountainous and hilly terrain, presenting a landscape pattern often described as “70% mountains, 20% water, and 10% farmland.” The area features a dense river network and abundant precipitation concentrated during the typhoon and plum rain seasons, which inherently confers a high potential erosion risk. Socio-economically, as a major birthplace of China’s private economy, Wenzhou has undergone rapid industrialization and urbanization since the beginning of the 21st century^5,6. Large-scale infrastructure construction, frequent land-use conversions, and intensive agricultural and forestry development have profoundly altered the surface landscape structure and ecological processes. The superposition of natural erosion potential and anthropogenic disturbance has resulted in complex soil erosion characteristics in this region. However, compared to extensively studied inland ecologically fragile areas such as the Loess Plateau and the Tibetan Plateau^7–10, this type of coastal hilly region with intense human-land interactions has not yet received sufficient attention.

Existing global and national-scale research on soil erosion has laid a solid theoretical foundation for erosion assessment and ecological protection, achieving significant progress in both methodological development and mechanism exploration. In terms of quantitative assessment and dynamic monitoring, the Revised Universal Soil Loss Equation (RUSLE)^11–14, its China-adapted version, the Chinese Soil Loss Equation (CSLE)¹⁵, and comprehensive ecosystem service models like InVEST have become mainstream technical methods for simulating the spatiotemporal dynamics of soil erosion and have been widely applied in inland ecologically fragile areas¹⁶. Although these mainstream models provide powerful tools for the quantitative assessment of soil erosion, in-depth analysis using multivariate statistics and machine learning methods is still required to uncover the complex driving mechanisms behind it. Existing research typically combines geostatistical and attribution models to differentiate the impacts of natural and human factors on erosion patterns^16–20. For example, the Geodetector has been widely used to quantify the explanatory power of various factors. Research in the Panxi region indicates that slope is the primary natural factor determining the spatial differentiation of erosion, while land-use change has become the key anthropogenic driver of pattern evolution²¹. In the karst area of Anshun City, the Random Forest model identified vegetation cover and management factors, soil and water conservation practice factors, and rocky desertification intensity as the main drivers, revealing that interactions between factors often exhibit nonlinear enhancement effects²². Spatiotemporal correlation analysis in a typical watershed on the Loess Plateau further pointed out that increased extreme rainfall events may offset the erosion reduction benefits brought by vegetation restoration, highlighting the importance of the feedback mechanisms between climate change and human activities²³. These studies generally confirm that although natural conditions such as topography and precipitation establish the spatial baseline for erosion, anthropogenic interventions like land-use change are increasingly becoming the dominant force shaping the spatiotemporal dynamics of erosion^3,24. However, existing driver analyses often focus on single statistical methods or static comparisons of factor contributions^25,26, and remain insufficient in addressing the nonlinear relationships of complex multi-factor interactions, their spatiotemporal heterogeneity, and dynamic changes in their evolutionary processes. For coastal hilly regions like Wenzhou, where natural and anthropogenic pressures are highly intertwined and rapidly changing, more effective interpretable machine learning methods are urgently needed to more precisely clarify the driving mechanisms of soil erosion evolution.

Therefore, this study takes Wenzhou City as the research area. Based on multi-source remote sensing and meteorological data from 2000 to 2023, the InVEST model is employed to assess spatiotemporal changes in soil erosion intensity. Furthermore, hotspot analysis, Pearson correlation analysis, and the XGBoost-SHAP machine learning method are used to reveal the influence of natural and anthropogenic factors on the spatiotemporal evolution patterns of soil erosion. The findings are expected to provide precise scientific support for ecological conservation and regional spatial governance in Wenzhou and similar areas, while also contributing case-specific insights for deepening research on human-land system coupling.

Materials and methods

Study area

Located in southeastern Zhejiang Province, Wenzhou is bordered by the East China Sea to the east. Its geographic coordinates range from 27°03′-28°36′N and 119°37′-121°18′E, with a total land area of 12,110 km². The terrain is characterized by a pattern of “mountains on three sides and the sea on one,” with higher elevations in the southwest and lower in the northeast. From west to east, the landscape transitions sequentially through medium-low mountain areas, low hilly basins, plain and tidal flat zones, and coastal island regions (Fig. 1).

Fig. 1.

Overview map of the research area.

The climate falls within the central subtropical monsoon zone, marked by distinct seasonal shifts between winter and summer monsoons and four clearly defined seasons. The average annual temperature ranges from 17.3 °C to 20.0 °C, with annual precipitation between 1,023 mm and 2,494 mm. The region experiences frequent plum rains during the late spring and early summer, and is susceptible to tropical cyclones from July to September, contributing to a generally humid and mild climate.

The hydrological system is dominated by three major river networks—the Oujiang, Feiyunjiang, and Aojiang—complemented by the eastern plain river network. Soil types follow zonal distribution patterns: red soil and yellow soil prevail in the western mountainous areas, while paddy soil and fluvo-aquic soil are common in the eastern plains. Coastal tidal flats feature saline-alkali soil, and soil density is generally higher on the south bank of the Oujiang River than on the north bank, reflecting pronounced spatial differentiation.

Data sources and processing

Topographic and hydrological parameter data

The Digital Elevation Model (DEM) data used in this study has a spatial resolution of 30 m and was sourced from the Geospatial Data Cloud platform. Acquired via spaceborne remote sensing technology, this dataset is characterized by high precision and spatial continuity, enabling it to accurately reflect the topographic relief features of Wenzhou City. The data processing procedure was as follows: First, the original DEM data was clipped using ArcGIS PRO software to extract valid data within the study area boundary. Subsequently, hydrological analysis tools, including filling sinks, flow direction calculation, and flow accumulation analysis, were employed to derive topographic factors (such as slope, aspect, and topographic relief) and hydrological parameters (such as surface runoff paths and catchment areas).

Land use data

Land use data were derived from the China Multi-period Land Use Remote Sensing Monitoring Dataset (CNLUCC), provided by the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences (Resource and Environmental Science and Data Center). To ensure temporal consistency across datasets, this study adopted a time-node analysis strategy based on the availability of Land Use/Cover Change (LUCC) data. The dataset is provided in TIFF raster format boasting a spatial resolution of 30 m, with a temporal span covering 2000, 2005, 2010, 2015, 2020, and 2023. Relying primarily on Landsat series remote sensing imagery, this dataset was generated via manual visual interpretation and employs a two-tier classification scheme²⁷. The Level I categories comprise six classes: cultivated land, forest land, grassland, water bodies, construction land, and unused land. Regarding data processing, the raw raster data were initially extracted using the Wenzhou City administrative boundary vector to isolate the land use information for the study area. Subsequently, a projection transformation was performed on the clipped data to ensure a consistent spatial reference with ancillary datasets. Finally, based on the dataset’s validation report, the 30 m product demonstrates an overall classification accuracy surpassing 85% and a Kappa coefficient above 0.8, thereby fulfilling the accuracy prerequisites of this research.

Normalized difference vegetation index (NDVI) data

The NDVI dataset is composed of two distinct segments: the series spanning 2000 to 2020 was derived from Landsat TM imagery retrieved from the International Scientific Data Service Platform, whereas the 2023 records employed the MODIS NDVI product (MOD13Q1), featuring a spatial resolution of 250 m and a temporal interval of 16 days²⁸.

The processing steps were as follows: First, the TM imagery-derived NDVI data were clipped and projected to maintain consistent spatial reference with other data. Second, In order to meet the input data resolution requirements of the InVEST model and ensure its proper operation the MODIS NDVI data were mosaicked and resampled to unify the spatial resolution to 30 m, ensuring consistency in data scale. Ultimately, the Maximum Value Composite (MVC) technique was utilized to process the monthly NDVI records for every year to derive annual peak NDVI values, which reflect the prime condition of vegetation cover within the investigated region.

Meteorological data

Meteorological data were obtained from the CRU TS monthly dataset produced by the UK National Centre for Atmospheric Science (NCAS). This dataset covers global surface climate variables, features a complete time series and good spatial consistency, making it suitable for regional-scale climate characteristic analysis. This study selected key meteorological elements including precipitation, air temperature, and potential evapotranspiration, with a time span from 2000 to 2023. The CRU TS dataset, with a native spatial resolution of 0.5°×0.5°, was utilized to capture long-term climatic trends1. We employed Inverse Distance Weighting (IDW) to downscale this data to the study resolution. We acknowledge that this method may smooth out micro-climatic variations caused by Wenzhou’s complex topography (e.g., orographic precipitation). However, previous studies indicate that for regional-scale erosion modeling, the spatial heterogeneity of soil loss is predominantly controlled by the LS factor (derived from 30 m DEM) and land use, while climatic data primarily determines the magnitude of erosion potential. Thus, the resolution of meteorological data is considered acceptable for the objectives of this study.

The processing involved first extracting the meteorological grid point data corresponding to the study area and converting the monthly data into annual data. Following this, the Inverse Distance Weighting (IDW) interpolation technique was employed to conduct spatial interpolation on the meteorological data, yielding meteorological raster layers at a 30 m spatial resolution to ensure spatial consistency with ancillary datasets.

Soil moisture data

The soil moisture data used in this study are the annual Wetness (WET) data. The source data is the MODIS product MOD09A1 released by the National Aeronautics and Space Administration (NASA), which provides multi-band land surface reflectance observation data. The data coverage spans from 2000 to 2023 with a spatial resolution of 500 m. The coordinate system was uniformly converted to GCS_WGS_1984, maintaining consistency with the spatial reference of other topographic and meteorological data in the study area, thereby ensuring data spatial compatibility and analytical applicability.

Soil type distribution data

The soil-related data were obtained from the 1:1 million World Soil Database (Harmonized World Soil Database, version 1.1) (HWSD) dataset (resolution of 1 km × 1 km) and the World Soil Database (Harmonized World Soil Database, version 1.1) (HWSD) compiled by the Food and Agriculture Organization (FAO) and the International Institute for Applied Systems Research (IIASA) in Vienna. The spatial distribution data of soil types in Wenzhou City were derived through cropping and projection.

Research methodology.

InVEST model

This study applied the InVEST model to compute the soil erosion modulus of sloping farmland in the target area. The Sediment Delivery Ratio (SDR) module of this model successfully simulates the spatial configuration of soil erosion by combining topographic features and sediment transport mechanisms. In terms of technical execution, the module employs the Digital Elevation Model (DEM) as the primary data to map the spatial differentiation features of soil erosion intensity within the watershed, subsequently measuring the sediment yield of each unit using raster operations. This is integrated with parameters such as terrain connectivity and surface runoff pathways to methodically estimate the sediment delivery ratio^29,30. Theoretically, the model is based on the Revised Universal Soil Loss Equation (RUSLE). This equation creates a pixel-scale soil loss estimation approach by incorporating six major erosion-influencing factors³⁰. The specific formula is displayed in Eq. (1), and the computation methods for each factor are listed in EqS. (2)–(8) respectively.

1

2

3

4

5

6

7

8

The Rainfall Erosivity (R-factor) serves as a crucial parameter for quantifying the capacity of rainfall to induce soil erosion, where its magnitude demonstrates a positive correlation with potential erosion hazards^31–33. To compute the R-factor, this research adopted the daily precipitation estimation framework developed by Xie et al.³². As illustrated in Eq. (2), R denotes the daily rainfall erosivity (unit: MJ mm hm⁻² h⁻¹ a⁻¹), whereassignifies the mean rainfall volume for days experiencing daily rainfall ≥ 10 mm (unit: mm), and stands for the model coefficient.

The Soil Erodibility (K-factor) (unit: t hm² h MJ⁻¹ mm⁻¹) quantifies the sensitivity of soil to separation and displacement driven by erosive agents. This parameter was derived via the EPIC model and subsequently refined employing the correction equation introduced by Zhang et al.³⁴.

The Slope Length and Steepness (LS) factor embodies the joint impact of terrain characteristics on soil erosion. To be precise, the L-factor was calculated following the approach introduced by Wischmeier and Smith³¹, whereas the S-factor was computed based on the slope equation developed by McCool et al.³⁵. Regarding the variable definitions, denotes the slope length (unit: m), corresponds to the slope length exponent, and signifies the slope angle.

The Cover and Management factor (C-factor) serves as a vital premise for gauging the capacity of vegetation to suppress soil erosion, acting as a pivotal component in the precise estimation of soil loss and the development of soil and water conservation tactics. In order to capture the spatial heterogeneity of vegetation cover, this research calculated this factor relying on the vegetation coverage fraction³⁶. Within Eqs. (6) and (7), C represents the cover and management factor, f signifies the vegetation coverage fraction, NDVI_soil stands for the NDVI value corresponding to bare soil or areas devoid of vegetation, and NDVI_max denotes the peak NDVI value observed under full vegetation cover conditions.

The Support Practice (P-factor) quantifies the efficacy of engineering and management interventions in mitigating soil erosion. At present, a universally accepted protocol for determining the P-factor in broad-scale research is lacking. In order to rationally depict its spatial configuration, this study utilized the empirical equations suggested by Lufafa et al.³⁷ and Teng et al.³⁸ for computation. Within Eq. (8), signifies the support practice factor, whiledenotes the slope percentage (%).

Hot spot analysis (Getis-Ord Gi*)

Hot spot analysis, grounded in spatial autocorrelation principles and the Getis-Ord Gi* statistic, serves as a traditional approach for detecting spatial agglomeration features of geographic components. By computing local statistics regarding attribute values among a specific spatial unit and its adjacencies, this technique integrates significance tests of z-scores and p-values to quantitatively ascertain the spatial grouping patterns of high and low magnitudes. Within the context of this research, this methodology is applied to examine the spatial clustering attributes of soil erosion intensity across Wenzhou City. To be precise, regions exhibiting statistically significant high-value agglomeration are classified as erosion hot spots, whereas those with significant low-value clustering are termed erosion cold spots; conversely, areas lacking significant clustering traits are designated as non-significant zones. Such a strategy aids in uncovering the spatial clustering distributional patterns of soil erosion intensity in Wenzhou City.

XGBoost regression model

In this study, the Extreme Gradient Boosting (XGBoost) algorithm was employed to model the non-linear relationship between environmental drivers and soil erosion³⁹. XGBoost is an optimized distributed gradient boosting library designed to be highly efficient, flexible, and portable⁴⁰. Unlike traditional Gradient Boosting Decision Trees (GBDT), XGBoost incorporates a regularization term in its objective function to control model complexity and prevent overfitting, which is critical for datasets with complex interactions.

For a dataset with n samples and m features, where represents the environmental factors (e.g., precipitation, slope) and represents the soil erosion modulus, the predicted value is the sum of the scores predicted by additive functions:

9

where is the space of regression trees (CART). To learn the set of functions , XGBoost minimizes the following regularized objective function:

10

Here, is a differentiable convex loss function (e.g., Mean Squared Error) that measures the difference between the prediction and the target . The term penalizes the complexity of the model to avoid overfitting, defined as:

11

where is the number of leaves in the tree, represents the leaf scores, and γ and λ are constants controlling the degree of regularization. This regularization strategy allows XGBoost to achieve better generalization performance compared to standard gradient boosting methods.

SHAP (SHapley Additive exPlanations) method

To interpret the “black-box” nature of the XGBoost model and quantify the contribution of each environmental factor, we utilized the SHAP method based on cooperative game theory⁴¹. SHAP assigns each feature an importance value for a particular prediction⁴⁰.

The SHAP value for a specific feature is calculated as the weighted average of its marginal contributions across all possible feature combinations⁴². The formula is expressed as:

12

where is the set of all features, is any subset of excluding feature , and is the model’s prediction for instance using only the features in . The term represents the marginal contribution of feature in the context of subset .

Global Feature Importance: To assess the overall importance of each driver across the entire study area (as shown in the Feature Importance bar plots), we calculated the mean absolute SHAP value for each feature :

13

where is the total number of samples. A higher indicates that feature has a stronger influence on soil erosion.

SHAP Dependence Analysis: Furthermore, SHAP dependence plots were generated to visualize the functional relationship between specific feature values (e.g., precipitation amount) and their corresponding SHAP values. This allows for the identification of non-linear thresholds and response patterns where environmental factors either exacerbate (positive SHAP value) or mitigate (negative SHAP value) soil erosion.

Predictive modeling strategy and accuracy validation

In this study, the XGBoost algorithm was employed to predict soil erosion. Unlike distance-based algorithms (e.g., k-NN), XGBoost utilizes a tree-based ensemble approach. Therefore, feature normalization or standardization was not applied, as the algorithm’s split points are determined by ranking feature values, making it robust to variations in feature scales.The model performance was evaluated using the Coefficient of Determination (R²) and Root Mean Square Error (RMSE). For the datasets spanning 2000–2023, the model achieved robust accuracy (Average R² > 0.83, Average RMSE < 0.15), providing a solid foundation for the subsequent SHAP interpretability analysis.

Pearson correlation analysis and SHAP feature contribution analysis

To investigate the linear correlation characteristics between influencing factors and soil erosion intensity in Wenzhou City, and to clarify the relative importance and direction of effect of each driving factor on erosion intensity, this study combines Pearson correlation analysis and SHAP feature contribution analysis.

Pearson correlation analysis, based on statistical significance testing, quantifies the degree of linear correlation between erosion influencing factors and erosion intensity. It intuitively reflects the closeness of their association and provides a statistical basis for screening core influencing factors.

SHAP feature contribution analysis is grounded in the Shapley value from cooperative game theory. By constructing an additive feature attribution model, it quantifies the marginal contribution of each influencing factor to the prediction result of erosion intensity. This method can reveal both the global importance of each factor and the positive or negative driving effect of a single factor on the erosion intensity in local areas. It effectively compensates for the limitation of traditional feature importance analysis, which struggles to explain the mechanism behind factor effects.

All spatial analysis and visualization in this study were performed using ArcGIS Pro 3.3.2 (https://pro.arcgis.com), with the InVEST model version 3.17.0 (https://naturalcapitalproject.stanford.edu/software/invest), and Python 3.12 was employed for the primary modeling and analysis.

Result

Soil erosion characteristics

Temporal variation characteristics of soil erosion intensity

All analyses presented are based on the six time nodes: 2000, 2005, 2010, 2015, 2020 and 2023. Meanwhile, in this study, the degree of soil erosion was classified based on the soil erosion classification standards promulgated by the Ministry of Water Resources of the People’s Republic of China⁴³. From 2000 to 2023, the overall soil erosion intensity in Wenzhou City exhibited a trend of “initial minor fluctuations with slight strengthening, followed by continuous weakening,” and was consistently dominated by Slight and Light erosion levels (Table 1).

Table 1.

Classification characteristics of soil erosion intensity.

Different periods	Soil erosion level	Slight	Light	Middle	Strong	Very strong	Severe
2000	Area (km2)	8593.07	2388.56	3.96	0.42	0.08
	Percentage (%)	78.22	21.74	0.04	0	0
2005	Area (km2)	8560.19	2419.06	6.27	0.44	0.12
	Percentage (%)	77.92	22.02	0.06	0	0
2010	Area (km2)	8462.85	2494.29	27.91	0.64	0.39	0.003
	Percentage (%)	77.04	22.7	0.25	0.01	0	0
2015	Area (km2)	8528.97	2444.45	11.3	0.48	0.26
	Percentage (%)	77.63	22.25	0.1	0	0
2020	Area (km2)	8708.41	2276.24	1.12	0.31	0.01
	Percentage (%)	79.27	20.72	0.01	0	0
2023	Area (km2)	8627.78	2355.1	2.61	0.39	0.06
	Percentage (%)	78.53	21.44	0.02	0	0

In 2000, the area of Slight erosion was 8,593.07 km², accounting for 78.22% of the total regional area; the Light erosion area was 2,388.56 km², accounting for 21.74%; while the area of Middle and higher intensity erosion was only 4.46 km², constituting less than 0.1% of the total. This indicated that regional erosion was generally at a low intensity level. Between 2000 and 2010, the Slight erosion area decreased slightly to 8,462.85 km², while the Light erosion area increased to 2,494.29 km². Concurrently, the Middle erosion area increased from 3.96 km² to 27.91 km², and patches of Very Strong and Severe erosion appeared for the first time, reflecting a brief period of intensified erosion during this stage.

After 2010, regional erosion intensity entered a phase of continuous weakening: by 2015, the Slight erosion area had rebounded to 8,528.97 km², the Light erosion area had decreased to 2,444.45 km², and the area of Middle and higher intensity erosion had shrunk to 11.3 km². In 2020, the Slight erosion area further increased to 8,708.41 km², the Light erosion area dropped to 2,276.24 km², and the area of Middle and higher intensity erosion was only 1.43 km². By 2023, the areas of Slight and Light erosion were 8,627.78 km² and 2,355.1 km² respectively, while the area of Middle and higher intensity erosion was merely 2.61 km², with Severe erosion patches having completely disappeared. This demonstrates the significant effectiveness of regional ecological restoration measures.

Spatial variation characteristics of soil erosion intensity

The spatial pattern of soil erosion in Wenzhou City has long been characterized by “Slight erosion in the west, Light erosion in the east, and sporadic distribution of high-intensity erosion,” with spatial differences gradually diminishing over time (Fig. 2).

Fig. 2.

Spatial distribution characteristics of soil erosion intensity.

Regarding the spatial differentiation of erosion types, Slight erosion has long been concentrated in the low hill and plain areas of the western and southern parts of the study area. This region features gentle topographic relief and high vegetation coverage, with land use primarily consisting of forest land and farmland, resulting in weak hydraulic erosion dynamics and consequently maintaining a state of Slight erosion. Light erosion primarily covers the eastern coastal hilly areas. This region has relatively greater slope gradients and is influenced by coastal winds and seasonal rainfall, leading to a slightly higher erosion intensity compared to the western plains, but overall remaining at a Light level.

The spatial distribution of Middle and higher intensity erosion exhibits a distinct “phased and sporadic characteristic.” From 2000 to 2005, high-intensity erosion occurred only as very small patches distributed on local steep slopes within the eastern hills. By 2010, the number of high-intensity erosion patches had slightly increased, concentrated in small mountainous areas along the northeastern coast, likely associated with local vegetation destruction and engineering activities. After 2015, high-intensity erosion patches rapidly diminished. Between 2020 and 2023, there were no significant concentrated distribution areas of high-intensity erosion; only very small areas of Middle erosion persisted in individual bare rock areas with slopes > 25°, reflecting a significant reduction in the spatial aggregation of regional erosion.

Spatiotemporal transfer characteristics of soil erosion intensity

From 2000 to 2023, the spatiotemporal transfer of soil erosion in Wenzhou City was primarily characterized by “minor conversions between low-intensity erosion types,” while the transfer of high-intensity erosion featured “rapid expansion followed by decline (Fig. 3).”

Fig. 3.

Transfer of soil erosion intensity in Wenzhou City from 2000 to 2023.

In terms of the direction of type transfers, the mutual transfer between Slight and Light erosion was the core process. Between 2000 and 2010, approximately 130.22 km² of area transferred from Slight to Light erosion, mainly occurring in the transition zone between Slight and Light erosion in the eastern hills, primarily due to topographic and rainfall factors leading to increased erosion intensity. From 2010 to 2023, an area of 219.19 km² transferred from Light to Slight erosion. The areas undergoing this transfer highly overlapped with the areas that had previously changed from Slight to Light erosion, demonstrating the reverse regulatory effect of ecological restoration on erosion intensity.

The transfer process for Middle and higher intensity erosion was characterized by “short-term expansion and long-term contraction”: From 2000 to 2010, approximately 21.64 km² of area transferred from Light to Middle erosion, alongside stepwise transfers from Middle to Strong and then to Very Strong erosion (with transfer areas of 0.22 km² and 0.003 km², respectively), forming an “expansionary transfer” pattern for high-intensity erosion. From 2010 to 2023, a pattern of reverse transfer for high-intensity erosion became dominant. An area of 25.30 km² transferred from Middle to Light erosion, while all areas of Strong and Very Strong erosion transferred to Middle or lower intensity levels, and Severe erosion patches disappeared entirely. This reflects a “regressive transfer” pattern for high-intensity erosion.

In general, the intensity of spatiotemporal transitions in soil erosion within Wenzhou City between 2000 and 2023 exhibited a trajectory of “initially rising followed by a decline.” Subsequent to 2010, the prevailing transfer direction transitioned toward a “shift from high intensity to low intensity,” signifying a progressive enhancement in the stability of regional erosion.

Spatial clustering characteristics of soil erosion intensity

Findings derived from hot spot analysis reveal that the soil erosion intensity in Wenzhou City demonstrates pronounced spatial agglomeration features, characterized by a distinct geographical heterogeneity pattern in the regional configuration of high and low values (Fig. 4).

Fig. 4.

Analysis of soil erosion intensity cold and hot spots.

Hot spot areas, representing clusters of high erosion intensity, are concentrated in the northeastern, central, and southern regions. These areas are characterized by a core of most significant hot spots with strong spatial continuity. They are bordered by ribbon-like zones of non-significant areas, reflecting a gradient transition feature where erosion intensity decreases from high to low. Cold spot areas, representing clusters of low erosion intensity, are stably distributed in the western region. Centered around most significant cold spots, their spatial extent and boundary characteristics remain consistent across the years, with only weak connections to non-significant areas at the edges.

From a temporal evolution perspective, the most significant clustering extent of hot spot areas shows a characteristic of “first expanding and then contracting.” The coverage continuously expanded from 2000 to 2010 and gradually contracted from 2010 to 2023. Furthermore, the distribution of hot spot areas at the 95% and 90% confidence levels tended to become more fragmented. In contrast, the most significant clustering extent of cold spot areas remained stable throughout, without noticeable expansion or contraction. The distribution range of non-significant areas gradually expanded over time, and the boundaries between cold and hot spot areas became progressively blurred, reflecting a reduction in the degree of spatial differentiation of soil erosion intensity.

Overall, the spatial clustering characteristics of soil erosion intensity in Wenzhou City are manifested as: hot spot areas concentrated in the northeast-central-south, cold spot areas stable in the west, and non-significant areas forming ribbon-like transitions. This pattern provides clear geographical guidance for the spatial regulation of regional soil erosion.

Characteristics of influencing factors

Characteristics of natural factors

The spatial patterns of natural elements within the research region are predominantly shaped by the synergistic interplay of terrain features and climatic factors. From 2000 to 2023, the overall pattern of each factor remained stable, with only minor interannual fluctuations observed locally.

The Digital Elevation Model (DEM) reveals that the mountainous areas in the western and northern regions, such as Taishun and Wencheng, are high-altitude zones, while the coastal areas and the Oujiang River estuary in the eastern and southern parts are low-altitude areas. The overall topographic pattern remained largely stable throughout the study period. A certain correlation exists between slope distribution and the DEM. The western and northern mountainous areas exhibit dense, dendritic patterns of high slopes on the slope map, realistically reflecting the complex combination of mountain valleys and slopes. In contrast, the eastern and southern coastal plains are predominantly characterized by extensive low slopes, consistent with plain topography (Fig. 5).

Fig. 5.

Terrain factor characteristics.

Precipitation distribution is characterized by higher levels in the western and northern mountainous areas and lower levels in the eastern and southern coastal regions. Between 2000 and 2023, the spatial gradient of precipitation change was relatively gentle. Although interannual precipitation fluctuated slightly in some local areas, the overall pattern did not undergo significant changes (Fig. 6). Temperature distribution is characterized by higher values in the coastal plains and lower values in the western and northern mountainous areas. This temperature distribution pattern remained stable during the study period, with only weak interannual variations in local temperatures, closely related to topographic elevation (Fig. 7). Evaporation is higher in the central and eastern coastal areas and lower in the western and southern mountainous regions. From 2000 to 2023, the extent of areas with high evaporation slightly decreased, but the overall distribution trend did not change significantly, aligning with regional heat conditions and land cover types (Fig. 8).

Fig. 6.

Characteristics of precipitation factor.

Fig. 7.

Characteristics of temperature factor.

Fig. 8.

Characteristics of evaporation factor.

Soil moisture is higher in coastal plains and river network areas and lower in inland mountainous regions. During the study period, the extent of high soil moisture areas slightly expanded, indicating minor adjustments in regional moisture conditions (Fig. 9). Its distribution is closely related to precipitation patterns, topographic hydrological characteristics, and coastal proximity.

Fig. 9.

Characteristics of soil moisture factor.

Characteristics of human activity factors

From 2000 to 2023, the land use structure in Wenzhou was predominantly composed of forest land and cultivated land, together accounting for over 90% of the total regional area. However, the area and proportion of various land use types exhibited significant temporal variation. As the dominant land type in the region, forest land area fluctuated and decreased from 8,714.53 km² in 2000 to 8,265.46 km² in 2023, with its proportion gradually declining from 76.27% to 72.34%. Over the 23-year period, the cumulative reduction reached 449.07 km², making it the land use type with the largest area reduction during the study period. Cultivated land area showed a fluctuating characteristic of “decreasing first and then increasing”: it decreased from 2,052.37 km² in 2000 to 1,816.86 km² in 2010, and then gradually rebounded after 2010 influenced by cultivated land protection policies and agricultural structural adjustments, reaching 2,093.71 km² in 2023, with its proportion slightly increasing from 17.96% to 18.32%.

Construction land was the most dramatically changing land use type during the study period. Its area continuously expanded from 419.33 km² in 2000 to 884.75 km² in 2023, with its proportion increasing from 3.67% to 7.74%. The area increased by 111.0% over the 23 years, intuitively reflecting the intensive urbanization process in Wenzhou. Grassland and water areas showed a continuous contraction trend. Grassland area decreased from 557 km² in 2000 to 173 km² in 2023, with its proportion dropping from 0.49% to 0.15%. Water area shrank from 2,337.3 km² to 1,797.3 km², with its proportion decreasing from 2.05% to 1.57%, reflecting the gradually intensifying effect of ecological land being squeezed by human activities. Unused land area always remained within a minimal range of 0.7 km² to 2.4 km², showing no significant change (Table 2).

Table 2.

Time variation characteristics of land use types.

Different periods	Land use type	Cultivated land	Forest land	Grassland	Water	Construction land	Unused land
2000	Area (km2)	2052.37	8714.53	5.57	233.73	419.33	0.07
	Percentage (%)	17.96	76.27	0.05	2.05	3.67	0
2005	Area (km2)	1983.13	8659.05	4.92	244.52	533.87	0.11
	Percentage (%)	17.36	75.79	0.04	2.14	4.67	0
2010	Area (km2)	1816.86	8709.95	3.73	235.78	659.14	0.16
	Percentage (%)	15.9	76.23	0.03	2.06	5.77	0
2015	Area (km2)	2046.15	8406.01	3.19	212.83	757.29	0.14
	Percentage (%)	17.91	73.57	0.03	1.86	6.63	0
2020	Area (km2)	2149.16	8256.54	1.88	191.27	826.58	0.17
	Percentage (%)	18.81	72.26	0.02	1.67	7.23	0
2023	Area (km2)	2093.71	8265.46	1.73	179.73	884.75	0.24
	Percentage (%)	18.32	72.34	0.02	1.57	7.74	0

From 2000 to 2023, the spatial pattern of land use in Wenzhou exhibited the core characteristics of “forest land as the dominant matrix, contiguous urban expansion, and fragmentation of ecological land.” As the regional ecological matrix, forest land was widely distributed in the western and northern mountainous and hilly areas, and its spatial distribution remained overall stable (Fig. 10). However, the connectivity of forest land patches gradually decreased around urban areas and along major transportation routes. Especially after 2010, forest land in the eastern plains was gradually encroached upon by cultivated land and construction land, showing a spatial form characterized by “fragmented patches and blurred boundaries.”

Fig. 10.

Spatial distribution characteristics of land use types.

Cultivated land was concentrated in the eastern coastal plains and river valleys. Its spatial changes exhibited a dual characteristic of “expansion at the periphery and reduction in the interior”: cultivated land patches increased in the peripheral low mountainous and hilly areas, represented by the low mountainous areas of Yongjia County in the north, while cultivated land in the core eastern urban areas was largely replaced by construction land, typified by Lucheng District and Longwan District, resulting in increasingly irregular patch shapes. The spatial expansion characteristic of construction land was the most significant. In 2000, it was mainly concentrated in a few urban nodes in the eastern coastal area. After 2005, it spread from core urban areas to the surroundings. From 2015 to 2023, contiguous urban built-up areas gradually formed, extending inland along the Oujiang River, the Wenruitang River, and the Shenhai Expressway, continuously increasing the spatial overlap with cultivated land and forest land.

The spatial distribution of grassland and water areas showed trends of contraction and fragmentation. Grassland patches were originally scattered along the edges of mountains, and by 2023, only small patches remained in Taishun County and Wencheng County in the north. Water areas were mainly composed of the Oujiang River, Feiyun River, and coastal wetlands. The spatial changes were reflected in water areas along river banks being occupied by construction land, while the spatial form of coastal wetlands remained overall stable. Overall, the spatial differentiation of land use in Wenzhou continued to intensify, and the squeezing effect of urban expansion on ecological and agricultural spaces became increasingly prominent.

The spatiotemporal transfer process of land use in Wenzhou from 2000 to 2023 exhibited significant phased differentiation and regularity, with the direction, intensity, and driving mechanisms of transfer showing diverse characteristics over time (Fig. 11).

Fig. 11.

Transfer of land use types in Wenzhou City from 2000 to 2023.

The period 2000–2010 was a rapid transition phase for land use, with the core transfers concentrated in two dominant directions. The cumulative transfer from forest land to cultivated land reached 233.55 km², and the total transfer from cultivated land to construction land reached 354.82 km². Concurrently, the transfer from water areas to construction land was 9.46 km², indicating the initial trend of ecological land being squeezed. During 2010–2015, the transfer pattern showed characteristics of bidirectional interaction. Transfers from forest land to cultivated land and from cultivated land to construction land were 339.42 km² and 1,683.15 km², respectively. Simultaneously, ecological restoration-type transfers from grassland to forest land and reverse transfers from construction land to cultivated land occurred, reflecting a synergistic characteristic of development and protection in the land use system transition. From 2015 to 2023, the transfer process tended to stabilize and coordinate, with ecology-oriented transfers significantly strengthening. The transfer from cultivated land to forest land reached 101.75 km², and from 2020 to 2023, newly added construction land was transferred to forest land for ecological restoration. During this period, transfers from forest land to cultivated land and from cultivated land to construction land were 361.58 km² and 3,848.63 km², respectively, with transfer intensity remaining relatively stable.

Regarding the overall transfer pattern, the cumulative transferred-out area of cultivated land during the study period exceeded 8,000 km², of which over 40% was converted to construction land. The cumulative transferred-out area of forest land exceeded 1,000 km², with nearly 70% converted to cultivated land. The cumulative transferred-in area of construction land exceeded 4,000 km², with over 85% originating from cultivated land. The cumulative transferred-out areas of ecological lands such as grassland and water areas were 3.84 km² and 19.76 km², respectively, indicating a relatively small scale of transfer.

The interactive effects of soil erosion and various factors

Soil erosion is the outcome of nonlinear coupling among multiple factors, with drivers such as topography, meteorological conditions, and underlying surface properties interacting to jointly shape the spatiotemporal patterns of regional soil erosion. Comprehensive analysis based on multi-period Pearson correlation matrices, SHAP feature importance ranking, and SHAP dependence and interaction plots indicates that the positive interaction between slope and precipitation is a key factor influencing soil erosion intensity (Figs. 12, 13, 14 and 15).

Fig. 12.

The interactive effects of soil erosion and various factors.

Fig. 13.

SHAP contribution of influencing factors.

Fig. 14.

Summary map of driving factors.

Fig. 15.

SHAP dependency graph.

The SHAP feature importance ranking reveals that slope has the most pronounced effect on soil erosion, with SHAP values ranging from − 0.4 to 0.6, indicating that steeper slopes consistently increase erosion risk. Precipitation is the second most influential factor, with higher precipitation amounts also corresponding to positive SHAP values and amplifying erosion intensity. The SHAP dependence plot for slope further demonstrates that as slope increases, its positive SHAP values rise significantly, particularly under conditions of low soil moisture, where the two factors exhibit a synergistic effect in enhancing erosion. The dependence plot for precipitation also shows that high precipitation combined with steep slopes exerts a stronger positive influence on erosion, confirming the interactive reinforcement between these two factors.

Correlation matrix results show that the correlation coefficient between slope and soil erosion remains consistently high, between 0.52 and 0.54. Although the correlation coefficient between precipitation and erosion fluctuates, the interaction between precipitation and slope significantly amplifies the erosion effect compared to that of either factor alone. These findings align with the dominant roles of slope and precipitation observed in the feature importance analysis, where the importance score for slope ranges from 0.455 to 0.477, and that for precipitation ranges from 0.262 to 0.291.

Soil moisture exhibits a negative regulatory role in the erosion process, reflecting the inhibitory function of the underlying surface on erosion. This mechanism is visually represented in its SHAP dependence plot, where high soil moisture corresponds to negative SHAP values, indicating its mitigating effect on erosion. In 2020, the feature importance of soil moisture exceeded that of land use type, which corresponds to the strengthening of its negative correlation coefficient with soil erosion during this period.

Furthermore, interactions among meteorological factors indirectly influence the erosion process. For instance, the correlation coefficient between evaporation and air temperature remains stable at approximately 0.91 over time. Their coordinated variation indirectly affects soil erodibility by regulating soil moisture, with high evaporation conditions reducing soil moisture and thereby increasing erosion risk to some extent.

In summary, regional soil erosion is governed by a multi-scale interactive system of factors, with the strength and direction of each factor’s influence dynamically adjusting over time, reflecting the complex response mechanisms of soil erosion processes to environmental changes.

Discussion

Evolution of soil erosion and influencing factors under ecological restoration

From 2000 to 2023, soil erosion intensity in Wenzhou City exhibited distinct phases. Before 2010, erosion intensity was generally high and fluctuated significantly; thereafter, it entered a period of sustained decline. This trend shift broadly coincides with the strengthening of soil and water conservation policies in Wenzhou and Zhejiang Province, including the integration of erosion control targets into local development plans and stricter regulatory oversight of construction projects^44–46. Although causality cannot be directly inferred from this study, the temporal alignment suggests that policy-driven land-use adjustments and ecological management measures may have contributed to the mitigation of regional soil erosion⁴⁷. Specific local implementation data further supports this linkage. According to statistical records, from 2009 to 2014, Wenzhou City completed a total of 7,324 hm² of cultivated land reclamation and consolidation. Through engineering measures, unused or inefficiently utilized land was converted into cultivable farmland, which optimized land use patterns and contributed to soil erosion improvement^48,49.

The influence weight of the soil moisture factor on the erosion pattern continued to rise after 2020, eventually surpassing land use type to become the dominant factor. This shift implies a transformation in the control mechanisms governing regional soil erosion, transitioning from early dominance by human activity-induced land-use changes towards a regulatory role played by soil moisture following ecosystem function improvement. Relevant research indicates that long-term vegetation restoration and slope management can effectively improve soil physicochemical properties, enhance soil infiltration capacity, reduce surface runoff generation, and ultimately lower erosion susceptibility through alterations in soil moisture conditions⁵⁰. Concurrently, increased soil moisture is closely linked to the accumulation of soil organic matter and the enhancement of aggregate stability. Litter accumulated during vegetation restoration is converted into organic matter, promoting the formation of water-stable aggregates, strengthening soil cohesion, and making the soil more resistant to raindrop splash and runoff scour⁵¹. Furthermore, improved soil moisture conditions can further influence hillslope hydrological connectivity, altering runoff pathways and sediment transport efficiency, thereby regulating the erosion process at a larger scale⁵². Consequently, the rising influence of the soil moisture factor is an indirect manifestation of improved soil erosion resistance resulting from ecological restoration, a process that requires long-term accumulation to manifest its effects.

Therefore, the mitigation of soil erosion in Wenzhou is the result of the combined effects of policy regulation and ecological restoration. The increased importance of the soil moisture factor can serve as a significant indicator of improved regional soil structure and ecological function. Future research could further validate the mechanism through which soil moisture affects erosion by measuring parameters such as soil infiltration rate and aggregate stability.

Limitations and future research directions

This research performed a comprehensive examination of the spatiotemporal dynamics of soil erosion in Wenzhou City; however, specific constraints exist and warrant attention in prospective research.

From a data perspective, this study primarily relied on remote sensing imagery and reanalysis datasets. Although suitable for macro-level monitoring, their spatial resolution is insufficient to capture micro-scale processes. For example, pixel-based land classification cannot distinguish structural differences in vegetation, and spatial heterogeneity in soil properties is not fully represented. These limitations may introduce bias in erosion estimates for critical areas such as steep slopes and active construction sites. Future research could integrate UAV aerial photography with ground-based sensor networks to establish an air-space-ground monitoring system. Such an approach would enable acquisition of higher-resolution field data—including vegetation cover and soil compaction—to validate the findings of this study at finer scales.

Regarding methodological approaches, the current research mainly focuses on the statistical relationships between erosion intensity and various factors. Although key factors like slope gradient and precipitation were identified, the understanding of complex non-linear interactions between these factors remains inadequate. Subsequent studies are recommended to incorporate process-based mechanistic models. Utilizing scenario simulations to quantitatively dissect the contribution of each driving factor would provide a deeper, process-oriented explanation from the perspectives of pedogenesis and eco-hydrology for phenomena such as the increased importance of soil moisture.

From a research perspective, the consideration of socio-economic factors needs strengthening. While this study analyzed land use as a direct manifestation of human activity, the exploration of underlying driving mechanisms, such as ecological compensation policies, industrial restructuring, and changes in rural labor forces, remains insufficient. These factors indirectly influence soil erosion processes by altering land management practices. Therefore, constructing a comprehensive assessment framework that integrates both physical geographical and socio-economic factors, and incorporating explanatory variables like policy intensity and management investment, would enable a more holistic revelation of the anthropogenic drivers behind soil erosion evolution.

Furthermore, this study primarily analyzed past and current erosion dynamics based on historical data. The potential intensifying impact of climate change, particularly the projected increase in frequency and intensity of extreme rainfall events, on future soil erosion risk in the region was not explicitly assessed. This represents an important limitation for long-term erosion forecasting and risk management. Future work should prioritize incorporating climate change projections. Utilizing downscaled climate model data (e.g., from CMIP6) to drive erosion models under different emission scenarios (e.g., SSP2-4.5, SSP5-8.5) would be a valuable extension. Such scenario-based projections would help quantify the relative contribution of climate change to future erosion patterns and inform the development of more resilient and adaptive soil conservation strategies.

Conclusion

This study systematically analyzed the spatio-temporal evolution of soil erosion and its interaction with driving factors in Wenzhou City from 2000 to 2023. The main conclusions are as follows:

1. Utilizing the InVEST model combined with time-series data analysis, it was determined that the overall soil erosion intensity in Wenzhou from 2000 to 2023 exhibited a phased characteristic: slight fluctuations with a modest increase before 2010, followed by a continuous and steady decline after 2010. Throughout the study period, the combined area of slight and light erosion consistently accounted for over 99% of the total, while the area of moderate and high-intensity erosion decreased from 27.91 km² in 2010 to 2.61 km² in 2023, with severe erosion patches completely disappearing. These findings indicate significant effectiveness of regional ecological restoration measures.

2. By applying hotspot analysis, the spatial characteristics of soil erosion in Wenzhou were clarified, showing a long-term pattern of slight erosion in the west, light erosion in the east, and scattered distribution of moderate to high-intensity erosion. Slight erosion is concentrated in the western low-mountain and plain areas, while light erosion covers the eastern coastal hilly regions. Erosion hotspots are primarily located in the northeast, central, and southern parts, whereas cold spots remain stable in the west. Both spatial heterogeneity and the clustering of high-intensity erosion have gradually decreased over time.

3. Employing Pearson correlation analysis and the XGBoost-SHAP machine learning method, it was established that soil erosion in Wenzhou results from the nonlinear coupling of natural factors and human activities. The correlation coefficient between slope gradient and soil erosion remained stable at 0.52–0.54, while that for precipitation ranged from 0.17 to 0.25, with their positive interaction identified as a core natural driving factor. Soil moisture showed a negative correlation with erosion, ranging from − 0.39 to -0.42, and its importance surpassed that of land use after 2020. The synergistic variation between evapotranspiration and temperature indirectly regulates soil erosion intensity by modulating soil water content.

Author contributions

H.H. conceived and designed the study, performed data collection and statistical analysis, interpreted the results, and wrote the main manuscript text. H.H. also prepared all tables and figures. The author reviewed and approved the final manuscript.

Funding

The author did not receive support from any organization for the submitted work.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.