Correlation Analysis of Geological Disaster Density and Soil and Water Conservation Prevention and Control Capacity: A Case Study of Guangdong Province

Abstract

This study investigates the spatial coupling between geohazard susceptibility and soil conservation capacity in Guangdong Province, China, using integrated spatial analysis and machine learning approaches. Through kernel density estimation, hotspot analysis, principal component analysis (PCA), and t-SNE clustering applied to 11,252 geohazard records and nine soil conservation factors, we identify three critical mechanisms: (1) Topographic steepness (LS factor) constitutes the primary control on geohazard distribution (r = 0.162, p < 0.001), with high-risk clusters concentrated in northeastern mountainous regions (Meizhou-Huizhou-Heyuan); (2) Vegetation coverage (C_mean) mediates rainfall impacts, exhibiting significant risk reduction (r = −0.099, p < 0.001) despite counterintuitive negative correlations with mean rainfall erosivity; (3) Soil conservation effectiveness depends on topographic context, reducing geohazard density in moderate slopes (Cluster 0: 527.04) but proving insufficient in extreme terrain (Cluster 2: LS = 20.587). The emerging role of rainfall variability (R_slope, r = 0.183) highlights climate change impacts.

1. Introduction

Geological hazards are destructive surface changes induced by geological processes or human activities, constituting natural disasters that pose severe threats to both society and ecosystems [1,2]. Global disaster inventories confirm that high-relief terrains and coastal zones are particularly prone to landslides, debris flows, and subsidence, events that frequently result in substantial socio-economic losses across these vulnerable regions. Critically, the escalating pace of climate change—manifested through increased frequency and intensity of extreme weather events—combined with the intensification of land-use practices, is dramatically amplifying the exposure and susceptibility of communities to these pervasive hazards [3]. This heightened risk profile underscores an urgent need for innovative mitigation strategies to enhance resilience [4,5].

Soil–water conservation (SWC) serves as a foundational disaster mitigation strategy by enhancing slope stability through erosion control [6]. Previous studies have demonstrated that vegetation restoration, structural stabilizers (e.g., lime, cement, and fly ash), and terrain engineering (e.g., soil compaction) can significantly mitigate rainfall-induced slope failures [7,8,9]. However, regionally divergent disaster susceptibilities persist, reflecting spatial heterogeneities in SWC implementation efficacy [10,11]. Thus, quantifying spatial linkages between geohazard density and SWC effectiveness emerges as a critical research frontier for sustainable risk governance [12].

Guangdong Province, situated in southern China, features diverse topography characterized by complex landforms and a humid climate with abundant rainfall. The region experiences various geological disasters, including landslides, collapses, and debris flows. Due to significant terrain undulations, intricate geological conditions, and high-intensity precipitation, Guangdong faces perennial threats from geological hazards. In recent years, accelerated urbanization and infrastructure development have intensified human impacts on the geological environment, leading to increased frequency and risk of geological disasters. To mitigate these risks, soil and water conservation measures have been widely implemented across Guangdong. These include vegetation restoration, soil improvement, and conservation engineering projects, which collectively enhance ecological conditions and soil structure [13]. For instance, the following may be considered: Vegetative cover and slope surface treatment reduce rainfall-induced soil erosion [14]; Barrier facilities decelerate debris flow velocity and dissipate destructive energy [15]. Such interventions not only lower disaster occurrence probabilities but also improve regional ecological resilience. However, the effectiveness of conservation measures varies significantly across regions, influenced by local geological settings, climatic patterns, and anthropogenic activities.

In recent years, scholars worldwide have made significant achievements in the research on geological hazards and soil–water conservation. Studies primarily focus on the following aspects: (1). Spatial distribution and influencing factors of geological hazards: Technologies such as remote sensing and Geographic Information Systems (GIS) are used to analyze spatial distribution characteristics and key influencing factors [16,17,18] (e.g., topography, rainfall, soil types) of geological hazards. These studies provide crucial data support for revealing the mechanisms of geological hazard occurrence. (2). Effectiveness evaluation of soil–water conservation measures [19,20]: Through field observations and experimental studies, the effectiveness of various soil–water conservation measures (e.g., vegetation restoration, engineering barriers) in reducing soil erosion and controlling geological hazards is evaluated. These studies provide a scientific basis for promoting and optimizing soil–water conservation practices. (3). Relationship between geological hazards and soil–water conservation [21,22]: Some scholars employ statistical analysis to explore the relationship between geological hazard frequency and soil–water conservation measures, finding that conservation factors (e.g., vegetation coverage [23], soil retention capacity [24]) significantly influence hazard occurrence. Despite these substantial accomplishments, most studies concentrate on single factors or localized regions, lacking systematic, global-scale spatial analysis of the relationship between geological hazard density and soil–water conservation capacity. Systematic research remains relatively scarce for regions like Guangdong Province, characterized by complex geological conditions and diverse hazard types. In particular, studies on the spatial distribution of geological hazard density and its correlation with soil–water conservation capacity are insufficient. The complexity of geological hazards and the diversity of conservation measures in Guangdong necessitate further research to uncover intrinsic relationships, providing scientific foundations for disaster prevention and mitigation.

This study aims to elucidate the spatial distribution characteristics of geological hazard density in Guangdong Province through spatial analysis methods and explore its relationship with soil and water conservation capacity. The specific objectives include: (1). Reveal spatial distribution patterns of geological hazard density: Utilize spatial hotspot analysis to uncover the spatial distribution laws of geological hazard density within the study area and identify high-risk zones. (2). Investigate spatial correlations between geological hazard density and soil and water conservation factor: Quantify the spatial correlation between geological hazard density and soil and water conservation factor, and identify key influencing factors. The findings will provide critical insights for geological hazard prevention and mitigation in Guangdong Province and other similar regions.

2. Materials and Methods

2.1. Study Area

Guangdong Province is located in the Pearl River Basin in southern China. The province features complex topography, primarily consisting of coastal plains, hills, and mountains. It experiences a humid subtropical monsoon climate, characterized by abundant annual rainfall. The average annual precipitation exceeds 1500 mm, with uneven distribution concentrated predominantly in the summer rainy season. Major rivers, including the Pearl River (Zhu Jiang), East River (Dong Jiang), and West River (Xi Jiang), form an extensive river network. This hydrological system exerts a crucial impact on regional soil–water conservation and the occurrence of geological hazards [25].

Guangdong Province exhibits a complex and diverse geological environment, primarily composed of granite, metamorphic rocks, and sedimentary rocks, representing distinct geological formations. The terrain features significant topographic relief, with mountainous areas accounting for over 76% of the total land area. Owing to complex topography and highly variable climatic conditions, the region is prone to frequent geological hazards, with landslides, rockfalls, and debris flows being the most prevalent types. During torrential rain seasons (April–September), intense rainfall and elevated soil saturation levels significantly increase the susceptibility to catastrophic geological events, particularly in mountainous and rural slope-cutting zones [26].

2.2. Data Source and Type

The spatial distribution data of geological hazard sites in Guangdong Province was obtained from the Resource and Environment Science and Data Center of the Chinese Academy of Sciences (https://www.resdc.cn/DataSearch.aspx) (accessed on 31 March 2024), encompassing 11,252 hazard sites across seven categories: collapses, sinkholes, debris flows, land subsidence, ground fissures, landslides, and unstable slopes [27].

The dataset of soil conservation service preventing soil water erosion in China (1992–2019) is available at Science Data Bank (https://cstr.cn/31253.11.sciencedb.07135) (accessed on 31 March 2024) [28]. This dataset includes nine zip files (“.rar”). All the data in the zip files are raster data (“.tif”). The details about every zip file we used were as follows (Figure 1).

“C1992–2019.rar”: It contains the mean value and changing rate of the C-factor from 1992 to 2019, which were named “C_mean.tif” and “C_slope.tif”, respectively.

“K_300.rar”: It contains the K-factor data, named “K_300.tif” (300 m × 300 m resolution).

“LS_300.rar”: It contains the LS-factor data, named “LS_300.tif” (300 m × 300 m resolution).

“P_300.rar”: It contains the P-factor data, named “P_300.tif” (300 m × 300 m resolution).

“R1992–2019.rar”: The mean value and changing rate of the R-factor in China from 1992 to 2019, named “R_mean.tif” and “R_slope.tif”, respectively. The resolution of the data is 1 km × 1 km.

“SC 1992–2019.rar”: The mean value and changing rate of soil conservation service (300 m × 300 m resolution).

Which, C is the vegetation cover and management factor (dimensionless); K is the soil erodibility factor (t h MJ⁻¹ mm⁻¹); LS is the topographic factor (dimensionless) with L being the slope length factor and S being the slope factor; P is the support practice factor (dimensionless); R is the rainfall erosivity factor (MJ mm ha⁻¹ h⁻¹ a⁻¹); SC is the soil conservation service (t ha⁻¹ a⁻¹) is defined as soil retention, which is soil water erosion prevented by vegetation and practice measures.

2.3. Method

(1) Using the Kernel Density Estimation (KDE) method [29,30], we generated a density distribution map of geohazard sites via ArcGIS Pro 3.0.0 to determine their spatial density distribution.

$$\mathrm{G}\mathrm{r}\mathrm{i}\mathrm{d}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}={\displaystyle \frac{1}{nh}}\sum _{i=1}^{n}K\left({\displaystyle \frac{x-x_i}{h}}\right)$$

(1)

where Gridcode represents the estimated density value at location x; n denotes the total number of geohazard sites; h is the bandwidth (smoothing parameter); K signifies the kernel function, with commonly used types including the Gaussian kernel and Uniform kernel; x_i indicates the location of the i-th geohazard site.

The bandwidth (search_radius) selection mechanism strictly follows Silverman’s rule of thumb for spatial correction algorithms. The tool automatically calculates the optimal radius based on the spatial distribution of the input points (11,252 disaster points) using the following formula:

$$h=0.9\times \mathit{min}\left(\widehat{\sigma }, {\displaystyle \frac{IQR}{1.34}}\right)\times n^{-{\displaystyle \frac{1}{5}}}$$

(2)

where $\widehat{\sigma }$ is the standard deviation of the nearest neighbor distances, IQR is the interquartile range, and $n$ = 11,252.

(2) Using ArcGIS Pro 3.0.0 and geohazard distribution data of Guangdong Province, we applied spatial statistical analysis to calculate the z-score and p-value for each geohazard site [31]. The z-score measures the local clustering intensity of geohazard sites, while the p-value (derived from the z-score based on the standard normal distribution) evaluates the statistical significance. Confidence intervals further determine the reliability range of the results. This approach ultimately identified significant hot spot and cold spot clusters within Guangdong Province, revealing the spatial distribution characteristics and their statistical significance of geohazards.

The steps to identify significant geological hazard clusters using the Z-score are as follows:

1. Calculate the local density: Compute the kernel density value $x_i$ for each location in Guangdong Province;

2. Calculate the global statistics:

$$\overline{x}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{x}_i$$

(3)

$$\mathrm{s}=\sqrt{{\displaystyle \frac{1}{n-1}}{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}$$

(4)

3. Calculate Z-score:

$$z_i={\displaystyle \frac{x_i-\overline{x}}{S}}$$

(5)

where x is the observed value; $\overline{x}$ is the average value.

4. Significance test for the Z-score of individual locations (Hotspot Analysis):

If $z_i\ge 2.58$, location $i$ is a significant hotspot at the 99% confidence level (i.e., the hazard density at this location is far above the average level);

If $1.96\le z_i<2.58$, location $i$ is a hotspot at the 95% confidence level;

If $-1.96\le z_i<1.96$, location $i$ is not significant (i.e., the hazard density is close to the average level);

If $-2.58 {\le z}_i<-1.96$, location $i$ is a significant cold spot at the 95% confidence level;

If $z_i\le -2.58$, location $i$ is a significant cold spot at the 99% confidence level (i.e., the hazard density at this location is far below the average level).

(3) By overlaying soil and water conservation indices with geohazard density distribution data, we calculated the correlations and significance levels between geohazard density and each factor. Principal Component Analysis (PCA) [32,33] was employed to extract dominant influencing factors and quantify their impact on geohazard density.

The input data includes 11,252 observation samples and nine soil and water conservation factor variables (C_mean, C_slope, K_300, LS_300, P_300, R_mean, R_slope, SC_mean, SC_slope). The analysis process is as follows:

1. Data Standardization: Use Z-standardization to eliminate the effects of dimensions:

$$Z=\sigma \left(x-\mu \right)$$

(6)

where $\mu$ is the mean of the variable, and $\sigma$ is the standard deviation.

2. Correlation Matrix Calculation: Construct a 9 × 9 Pearson correlation matrix $R$

$$\mathit{R}=\left[\begin{array}{ccc}{r}_{11}& \dots & r_{19}\\ ⋮& \ddots & ⋮\\ r_{91}& \dots & r_{99}\end{array}\right]$$

(7)

$$r_{ij}=\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{r}\left(x_i,x_j\right)={\displaystyle \frac{{\sum }_{k=1}^n\left(x_{ki-}\overline{x_i}\right)}{\sqrt{{\sum }_{k=1}^n{\left(x_{ki-}\overline{x_i}\right)}^2{\sum }_{k=1}^n{\left(x_{kj-}\overline{x_j}\right)}^2}}}$$

(8)

3. Eigen Decomposition: Solve for the eigenvalues and eigenvectors of the correlation matrix. The eigenvalues $\lambda$ satisfy the following:

$$\left|R-\lambda I\right|=0$$

(9)

where $R$ is the correlation matrix and $I$ is the identity matrix.

4. Component Selection: Retain the principal components based on the Kaiser criterion (eigenvalues > 1).

5. Factor Rotation: Apply Varimax rotation to optimize the structure of the loading matrix.

(4) $KMO$ Test: Conduct the Kaiser–Meyer–Olkin ($KMO$) test to quantitatively evaluate the sampling adequacy and variable intercorrelation strength, ensuring data suitability for factor analysis. The KMO calculation formula is as follows:

1. Calculate the partial correlation matrix:

$$P=-{\displaystyle \frac{{\mathit{R}}^{-1}}{\sqrt{diag\left({\mathit{R}}^{-1}\right)\times diag\left({\mathit{R}}^T\right)}}}$$

(10)

where, $R^{-1}$ is the inverse of matrix $R$ mathbit $R^T$ is the transpose of matrix ×$R$, and the diagonal elements are set to 0.

2. Calculate the $KMO$ value:

$$KMO={\displaystyle \frac{\sum _{i\ne j}{r}_{ij}^2}{\sum _{i\ne j}{r}_{ij}^2+\sum _{i\ne j}{p}_{ij}^2}}$$

(11)

(5) Bartlett’s Test of Sphericity: Perform Bartlett’s test to statistically verify the presence of significant systematic associations among variables by rejecting the null hypothesis of an identity correlation matrix (p < 0.01).

$${\chi }^2=\left[\left(n-1\right)-{\displaystyle \frac{2p+5}{6}}\right]\times \left(-\mathit{ln}\left|\mathit{R}\right|\right)$$

(12)

where $p$ represents the number of variables, which is 9. $\left|\mathit{R}\right|$ values are the determinants of the correlation matrix.

(6) t-Distributed Stochastic Neighbor Embedding (t-SNE) is a nonlinear dimensionality reduction method particularly suitable for high-dimensional data visualization. This method retains the local structure features of the data by optimizing the Kullback–Leibler divergence between the probability distributions in the low-dimensional and high-dimensional spaces. The principle is as follows:

1. High-Dimensional Space Similarity: For high-dimensional data points xi and xj, the conditional probability is defined as:

$$p_{j|i}={\displaystyle \frac{exp(-{‖x_i-x_j‖}^2/2{\sigma }_i^2)}{{\sum }_{k\ne i}exp(-{‖x_i-x_k‖}^2/2{\sigma }_i^2)}}$$

(13)

where ${\sigma }_i$ is the variance of the Gaussian distribution centered at $x_i$, controlled by the perplexity parameter.

$$Prep\left(P_i\right)=2^{H\left(P_i\right)}, H\left(P_i\right)=-{\sum }_j{p}_{j|i}{log}_2{p}_{j|i}$$

(14)

2. Low-Dimensional Space Similarity: In the low-dimensional embedding space, similarity qij is measured using a Student’s t-distribution:

$$q_{ij}={\displaystyle \frac{{(1+{‖y_i-y_j‖}^2)}^{-1}}{{\sum }_{k\ne l}{(1+{‖y_k-y_l‖}^2)}^{-1}}}$$

(15)

3. Optimization Objective Function: The objective is to minimize the Kullback–Leibler divergence between pairwise distributions:

$$KL(P\left.‖Q\right)={\sum }_i{\sum }_j{p}_{ij}log{\displaystyle \frac{p_{ij}}{q_{ij}}}$$

(16)

Using gradient descent optimization, the partial derivative with respect to $y_i$ is as follows:

$${\displaystyle \frac{\delta KL}{\delta y_i}}=4\sum (p_{ij}-q_{ij})(y_i-y_j){(1+{‖y_i-y_j‖}^2)}^{-1}$$

(17)

In this study, t-SNE is used to reduce the dimensionality of the standardized soil conservation factors (including eight indicators such as terrain ruggedness, vegetation cover index, and rainfall erosivity) from a high-dimensional space to a two-dimensional visualization space. The t-distribution characteristic of t-SNE effectively handles the long-tail distribution features present in this study (such as the slope distribution), avoiding the “crowding problem” and enhancing the visualization of geomorphological features. The parameters are set as follows:

Perplexity = 30, balancing local and global structures;

Learning rate = 200, ensuring the convergence of the optimization process;

Early exaggeration = 12, enhancing cluster separation;

Number of iterations (n_iter) = 1000, ensuring sufficient convergence of the algorithm.

(7) Subsequently, K-means clustering was employed to delineate characteristic zones using the following:

1. Core formula of K-means clustering:

Objective function:

$$J\left(k\right)={\sum }_{i=1}^k{\sum }_{x\in C_i}{‖x-{\mu }_i‖}^2$$

(18)

where $k$ is the number of clusters, $C_i$ is the set of samples in the $i$-th cluster, ${\mu }_i$ is the centroid of the $i$-cluster, and $‖x-{\mu }_i‖$ is the Euclidean distance from the sample point $x$ to the centroid ${\mu }_i$. The centroid ${\mu }_i$ is calculated as follows:

$${\mu }_i=\frac{1}{\left|C_i\right|}{\sum }_{x\in Ci}x$$

(19)

The Euclidean distance is calculated as follows:

$$d\left(x,y\right)=\sqrt{{\sum }_{j=1}^p{\left(x_j-y_j\right)}^2}$$

(20)

where $x,y$ are two p-dimensional sample points (where p = 9 soil and water conservation factors), and $x_j$, $y_j$ are the values of the $j$-th variable.

Cluster number selection criteria:

➀ Elbow method: The rate of decrease in $J\left(k\right)$ as $k$ increases:

$$∆\left(k\right)=\left|{\displaystyle \frac{J\left(k\right)-J\left(k+1\right)}{J\left(k+1\right)}}\right|$$

(21)

when $∆\left(k\right)$ is significantly smaller than $∆\left(k+1\right)$, choose $k$ as the optimal value.

2. Kruskal–Wallis test (overall difference test):

$$H=\left({\displaystyle \frac{12}{N\left(N+1\right)}}{\sum }_{i=1}^k{\displaystyle \frac{R_i^2}{n_i}}\right)-3\left(N+1\right)$$

(22)

where $N$ is the total sample size, $k$ is the number of clusters, $R_i$ is the rank sum of the $i$-th cluster, and $n_i$ is the sample size of the $i$-th cluster.

3. Mann–Whitney U test (pairwise comparison):

$$U=R_1-{\displaystyle \frac{n_1\left(n_1+1\right)}{2}}$$

(23)

where $R_1$ is the rank sum of the first group.

3. Results

3.1. Distribution Characteristics of Geohazard Density

Figure 2a displays the spatial distribution of geohazard types across Guangdong Province, indicating province-wide coverage with a particularly dense concentration of both variety and quantity in cities like Guangzhou and Shenzhen. Figure 2b illustrates regional geohazard intensity, where Areas (0.001–282.841) represent widespread low-intensity zones. Areas (1979.884–2825.644) denote concentrated high-intensity zones, primarily in Meizhou, Huizhou, Heyuan, and the surrounding regions. Figure 2c identifies geohazard hot spots and cold spots using hot spot analysis: Deep red areas (99% confidence level) cluster in Meizhou, Huizhou, and northern Heyuan. Orange areas (95% confidence level) show broader distribution, including peri-urban zones of Qingyuan City.

3.2. Correlation Between Geological Hazard Density and Soil Conservation Factors

Figure 3 illustrates the correspondence between kernel density values and nine erosion factors (C_mean, C_slope, K_300, LS_300, P_300, R_mean, R_slope, SC_mean, SC_slope). Correlation analysis demonstrates significant relationships between geological hazard density and soil–water conservation factors (p < 0.05) (Figure 4,

Table 1. Correlation Matrix.

	grid_code	C_mean	C_slope	K_300	LS_300	P_300	R_mean	R_slope	SC_mean	SC_slope
grid_code	1	−0.099	−0.018	0.019	0.162	0.007	−0.247	0.183	0.116	0.067
C_mean	−0.099	1	0.269	0.065	−0.549	−0.348	0.256	−0.237	−0.566	0.109
C_slope	−0.018	0.269	1	0.073	−0.137	−0.119	−0.059	0.002	−0.167	−0.126
K_300	0.019	0.065	0.073	1	−0.012	−0.154	−0.346	0.281	−0.047	0.023
LS_300	0.162	−0.549	−0.137	−0.012	1	0.258	−0.252	0.22	0.954	−0.262
P_300	0.007	−0.348	−0.119	−0.154	0.258	1	0.011	0.018	0.248	−0.098
R_mean	−0.247	0.256	−0.059	−0.346	−0.252	0.011	1	−0.244	−0.05	−0.345
R_slope	0.183	−0.237	0.002	0.281	0.22	0.018	−0.244	1	0.194	0.085
SC_mean	0.116	−0.566	−0.167	−0.047	0.954	0.248	−0.05	0.194	1	−0.422
SC_slope	0.067	0.109	−0.126	0.023	−0.262	−0.098	−0.345	0.085	−0.422	1

and

Table 2. Correlation and Significance Analysis Results.

Factor	Correlation	p-Value	Significance	Interpretation
R_mean	−0.247	0	***	p < 0.001
R_slope	0.183	0	***	p < 0.001
LS_300	0.162	0	***	p < 0.001
SC_mean	0.116	0	***	p < 0.001
C_mean	−0.099	0	***	p < 0.001
SC_slope	0.067	0	***	p < 0.001
K_300	0.019	0.0006	***	p < 0.001
C_slope	−0.018	0.0019	**	p < 0.01
P_300	0.007	0.2469	n.s.	p ≥ 0.10

Notes: Significance (p-Value): p-Value < 0.001, “***”; p-Value < 0.01, “**”; not significant, “n.s.”.

). Key correlations include: strong negative correlations with mean rainfall erosivity (r = −0.247) and mean vegetation coverage (C_mean = −0.099); strong positive correlations with topographic factor (LS_300 = 0.162), mean soil conservation capacity (SC_mean = 0.116), and rate of change in rainfall erosivity (R_slope = 0.183); while soil erodibility (K_300 = 0.019) and rate of change in soil conservation capacity (SC_slope = 0.067) exhibited slight yet statistically significant positive correlations.

3.3. Principal Components of Soil and Water Conservation Affecting Nuclear Density

The results of the factor analysis suitability test indicate that the Bartlett’s test of sphericity yielded a statistic of 153,644.1 (p < 0.001), strongly rejecting the null hypothesis of variable independence and supporting the suitability of the data for dimensionality reduction analysis (

Table 3. The results of the factor analysis suitability test.

Variable Pair	${\mathit{r}}_{\mathit{i}\mathit{j}}$	${\mathit{r}}_{\mathit{i}\mathit{j}}^2$	${\mathit{p}}_{\mathit{i}\mathit{j}}$	${\mathit{p}}_{\mathit{i}\mathit{j}}^2$
C_mean vs. C_slope	0.27	0.07	0.19	0.04
C_mean vs. K_300	0.06	0.00	0.19	0.04
C_mean vs. LS_300	−0.55	0.30	0.30	0.09
C_mean vs. P_300	−0.35	0.12	−0.29	0.09
C_mean vs. R_mean	0.26	0.07	0.41	0.17
C_mean vs. R_slope	−0.24	0.06	−0.13	0.02
C_mean vs. SC_mean	−0.57	0.32	−0.39	0.15
C_mean vs. SC_slope	0.11	0.01	−0.08	0.01
C_slope vs. K_300	0.07	0.01	−0.04	0.00
C_slope vs. LS_300	−0.14	0.02	0.11	0.01
C_slope vs. P_300	−0.12	0.01	−0.04	0.00
C_slope vs. R_mean	−0.06	0.00	−0.10	0.01
C_slope vs. R_slope	0.00	0.00	0.09	0.01
C_slope vs. SC_mean	−0.17	0.03	−0.13	0.02
C_slope vs. SC_slope	−0.13	0.02	−0.29	0.09
K_300 vs. LS_300	−0.01	0.00	−0.15	0.02
K_300 vs. P_300	−0.15	0.02	−0.07	0.00
K_300 vs. R_mean	−0.35	0.12	−0.38	0.14
K_300 vs. R_slope	0.28	0.08	0.26	0.07
K_300 vs. SC_mean	−0.05	0.00	0.12	0.02
K_300 vs. SC_slope	0.02	0.00	−0.12	0.01
LS_300 vs. P_300	0.26	0.07	0.20	0.04
LS_300 vs. R_mean	−0.25	0.06	−0.64	0.41
LS_300 vs. R_slope	0.22	0.05	−0.05	0.00
LS_300 vs. SC_mean	0.95	0.91	0.97	0.94
LS_300 vs. SC_slope	−0.26	0.07	0.39	0.15
P_300 vs. R_mean	0.01	0.00	0.16	0.02
P_300 vs. R_slope	0.02	0.00	−0.02	0.00
P_300 vs. SC_mean	0.25	0.06	−0.18	0.03
P_300 vs. SC_slope	−0.10	0.01	−0.09	0.01
R_mean vs. R_slope	−0.24	0.06	−0.04	0.00
R_mean vs. SC_mean	−0.05	0.00	0.61	0.37
R_mean vs. SC_slope	−0.35	0.12	−0.05	0.00
R_slope vs. SC_mean	0.19	0.04	0.08	0.01
R_slope vs. SC_slope	0.08	0.01	0.16	0.03
SC_mean vs. SC_slope	−0.42	0.18	−0.48	0.23
Total		2.90		3.25
KMO = 2.90/(2.90 + 3.25) = 0.47

and Figure 5). However, the KMO test value was 0.47, below the recommended threshold of 0.6, indicating weak commonality information among the soil and water conservation factors. Specifically, there is significant informational independence between variables, such as a strong direct association between LS_300 (topographical factor) and SC_mean (soil conservation capacity) (ρ = 0.97, ρ² = 0.94), whereas the association between vegetation cover (C_mean) and rainfall erosivity (R_mean) is mainly mediated by topography (r = −0.55 vs. ρ = −0.13). This complex interaction pattern reflects the spatial heterogeneity characteristics of the soil and water conservation system in Guangdong Province.

3.3.1. Results of Principal Component Analysis (PCA)

Principal Component Analysis (PCA) reveals that soil and water conservation factors explain 83.13% of the total variance across five principal components (Figure 6,

Table 4. Principal Component Loadings Table.

	PC1	PC2	PC3	PC4	PC5	PC6	PC7	PC8	PC9
C_mean	−0.462	0.067	0.233	−0.072	0.008	0.029	0.754	0.386	−0.067
C_slope	−0.158	−0.054	0.591	0.683	0.119	−0.234	−0.054	−0.291	−0.017
K_300	−0.01	−0.498	0.355	−0.237	−0.276	0.605	0.067	−0.35	0.022
LS_300	0.544	−0.014	0.116	−0.015	0.289	−0.041	0.383	−0.052	0.675
P_300	0.262	0.167	−0.304	0.588	−0.546	0.314	0.253	0.068	−0.028
R_mean	−0.134	0.596	0.09	−0.291	−0.378	−0.196	0.123	−0.566	0.135
R_slope	0.193	−0.428	0.127	−0.128	−0.565	−0.624	0.036	0.195	0.009
SC_mean	0.548	0.112	0.176	−0.122	0.183	−0.051	0.268	−0.14	−0.717
SC_slope	−0.201	−0.407	−0.557	0.105	0.18	−0.22	0.355	−0.509	−0.079

and

Table 5. Principal Component Variance Table.

Principal Component	Variance Explained	Cumulative Variance
PC1	0.314249	0.314249
PC2	0.192757	0.507006
PC3	0.13828	0.645286
PC4	0.098034	0.74332
PC5	0.087963	0.831283
PC6	0.079334	0.910616
PC7	0.052394	0.96301
PC8	0.035187	0.998197
PC9	0.001803	1

). PC1 (31.42% variance) has high positive loadings from topography (LS_300, 0.54) and soil conservation services (SC_mean, 0.55), and a high negative loading from vegetation cover (C_mean, −0.46). PC2 (19.28% variance) is driven positively by rainfall erosivity (R_mean, 0.60) and negatively by soil erodibility (K_300, −0.50) and soil conservation change rate (SC_slope, −0.41). Subsequent PCs capture complex interactions: PC3 (13.83%) shows inverse vegetation cover (C_slope, 0.59) and soil conservation change (SC_slope, −0.56); PC4 (9.80%) reflects synergy between engineering measures (P_300, 0.59) and vegetation change (C_slope, 0.68); PC5 (8.80%) indicates opposite responses in rainfall erosion (R_slope, −0.57) and engineering measures (P_300, −0.55).

3.3.2. t-SNE Dimensionality Reduction and Clustering Analysis

Cluster analysis identified five feature regions (contour coefficient = 0.246, Calinski–Harabasz index = 7046.5) (Figure 7,

Table 6. Clustering Feature Table.

Cluster	Disaster Density (Median)	Vegetation Coverage (Mean)	Terrain Factor (Mean)	Rainfall Erosivity (Average)	Soil Conservation (Mean)
0	527.04	0.197	3.339	9863.08	725.937
1	672.59	0.026	11.305	7730.72	2462.48
2	578.112	0.032	20.587	9617.99	5454.6
3	600.42	0.197	3.531	8542.3	781.424
4	101.835	0.206	1.306	10482.2	307.082

): Cluster 1 exhibits the highest disaster density (672.59), where low vegetation coverage (0.026) and steep terrain (topographic factor = 11.305) dominate risk factors, with moderate soil conservation capacity (2462.48) insufficient to offset topographic disadvantages. Cluster 2 shows extreme terrain (topographic factor = 20.587), driving high disaster density (578.11), where even the strongest soil conservation capacity (5454.60) remains inadequate. Clusters 3 and 0 exhibit similar vegetation coverage and terrain factors, but they differ in hazard levels (600.42 vs. 527.04). Cluster 3 is characterized by high soil conservation capacity (781.42), while Cluster 0 is influenced by high rainfall erosivity (9863.08). Cluster 4 represents the sole low-risk zone (disaster density = 101.84), where flat terrain (topographic factor = 1.306) and high vegetation coverage (0.206) synergistically mitigate the impact of maximum rainfall erosivity (10,482.20).

The spatial distribution of clusters (Figure 8) and corresponding geomorphic characteristics (Figure 9) reveal critical patterns: Clusters 1 and 3 share identical geomorphic composition (marine deposits: 76.3% + fluvial-lacustrine deposits: 11.8% + alluvial deposits: 15.8%) yet exhibit different disaster densities (672.59 vs. 600.42). Cluster 2 displays extreme geomorphic homogeneity (marine deposits: 93.5%), with the maximum topographic factor (20.587) confirming uncontrollable risks in open marine terrain. Cluster 0 contains the highest proportion of alluvial deposits (32.7%), driving its heightened sensitivity to rainfall erosion (9863.08) and resulting in moderate disaster density (527.04). Cluster 4 is enriched with gentle continental slopes (42.3% occurrence) and lacustrine plains, where flat terrain (topographic factor = 1.306) and dense vegetation (coverage = 0.206) completely suppress the strongest rainfall erosivity (10,482.20), achieving minimal disaster density (101.84).

4. Discussion

This study presents a comprehensive spatial analysis of geohazard density and soil conservation capacity in Guangdong Province. Our findings reveal fundamental relationships as follows:

Comprehensive analysis indicates that Guangdong Province exhibits diverse and widespread geohazard types, with higher intensity observed in eastern and northern regions (e.g., Meizhou and Qingyuan). Geohazard hotspots are primarily concentrated in areas surrounding Meizhou and Huizhou, while cold spots dominate cities such as Maoming and Yangjiang.

The correlation between the kernel density of geological disasters and soil and water conservation factors reveals two core mechanisms. (1) Natural protection effect: A high average rainfall (R_mean) represents a normal level, where a higher average rainfall promotes vegetation growth, and dense vegetation (C_mean) significantly inhibits geological disasters, confirming its ecological barrier function. This is consistent with existing research, which indicates that vegetation itself is beneficial for slope stability [34]. Additionally, for slopes affected by rainfall, the application of vegetation significantly increases the safety factor of the slope [35]. (2) Risk-driven pattern: Steep terrain (LS_300) directly triggers disasters, while the positive correlation of soil conservation (SC_mean) is an interesting phenomenon. This reflects that areas with stronger soil conservation capacity often have higher geological disaster risk levels (positive correlation). This positive correlation is the result rather than the cause [36]. It reveals that in high-risk areas (due to steep terrain), more management efforts (soil and water conservation projects) have been made, thereby increasing soil conservation. Therefore, high soil conservation is an indicator or marker of the strong management needs and actions in high-risk areas. In practice, we cannot use the level of soil conservation alone as an evaluation standard for geological disaster susceptibility. A comprehensive evaluation, including actual management measures, is needed.

Although the KMO value (0.47) is lower than the conventional threshold, the spatial heterogeneity of soil and water conservation factors it reflects is precisely the essential characteristic of the disaster system in Guangdong Province. A key warning is that the increasing trend of rainfall erosivity (R_slope) becomes an emerging source of risk, highlighting the potential threat of climate change. Extreme rainfall induces geological disasters, becoming a challenge for geological disaster prevention and control under climate change conditions [37,38]. The weak impact of soil characteristics (K_300) indicates that regional disaster differentiation is mainly dominated by terrain and climate. The five PCA dimensions form an integrated framework for soil–water conservation dynamics. PC1 and PC2 dominate (50.7% combined variance), revealing core drivers: PC1 highlights the conflict between topography-driven conservation efforts and inadequate vegetation management; PC2 underscores high rainfall erosivity coupled with degrading soil conservation capacity. This structure provides a quantitative basis for spatial differentiation of geological hazard risks, prioritizing regions where natural forces and conservation lapses intersect.

The partitioning results reveal three core patterns: (1) Absolute dominance of topography: Clusters 1–2 demonstrate that steep terrain is a sufficient condition for disasters, even with strong management measures (Cluster 2), the risk cannot be completely mitigated [39]; (2) Vegetation-topography synergistic barrier: Cluster 4 validates that flat terrain combined with dense vegetation can completely inhibit disasters, highlighting the combined necessity of landform modification and ecological restoration; (3) Dual-path mechanism in medium-risk areas: Cluster 0 and Cluster 3 demonstrate considerable similarity across multiple environmental indicators (e.g., vegetation coverage, terrain factor, soil conservation), but exhibit significant divergence in geomorphic composition. As the fundamental basis of surface materials, geomorphic composition governs the efficiency of water–soil interactions. The heightened erosion sensitivity in Cluster 0 arises from the triple hybrid of marine, fluvial-lacustrine, and alluvial deposits (98%), compounded by the amplification effect of arid deposits, leading to persistent erosion. Conversely, Cluster 3’s elevated disaster density is primarily attributed to the abrupt destabilization tendency inherent to marine-dominated sediments—despite lower erosion sensitivity, individual disaster events likely generate larger-scale impacts. The similarity in environmental parameters is superseded by geomorphic distinctions: under comparable vegetation and terrain conditions, disaster typology (gradual erosion versus catastrophic failure) more effectively explains disaster density variation than total erosion volume. This analysis substantiates the principle that geological foundations dictate disaster mechanisms. Future investigations should examine event-based rainfall records and historical disaster patterns (e.g., prevalence of large-scale landslides or liquefaction events in Cluster 3). Regarding mitigation strategies: Cluster 0 necessitates distributed interventions (e.g., vegetated swales); Cluster 3 requires prioritized monitoring of critical rainfall thresholds (e.g., installation of piezometers).

The geomorphology-disaster feedback reveals three universal mechanisms: (1) Absolute dominance of topography: In marine sedimentation zones (Clusters 1–2), disaster risk is inherently determined by coastal geomorphic configuration (e.g., cliff inclination, sediment consolidation), where engineering interventions (e.g., Cluster 2’s conservation capacity of 5454.60) offer only marginal risk reduction due to fundamental landscape constraints. (2) Hydrological sensitivity effect: The alluvial system (Cluster 0) exhibits amplified feedback to rainfall erosion dynamics, with runoff coefficient increases during storm events, mandating adaptive rain–flood management frameworks integrating real-time hydrological monitoring. (3) Geomorphology–ecology synergistic barrier: Slope transition zones (Cluster 4) develop efficient disaster buffering through geometric–ecological coupling—where gentle gradients and multilayer vegetation coverage (canopy + understory) jointly dissipate kinetic energy, establishing a replicable regional management paradigm for hazard mitigation efficacy.

5. Conclusions

Based on 11,252 geohazard records (Resource and Environment Science and Data Center) and multi-decadal soil conservation datasets (Science Data Bank, 1992–2019), this study employed Spatial Clustering (Kernel Density Estimation + Z-score hotspot analysis), Multivariate Statistics (PCA with KMO and Bartlett’s p < 0.01 validation), and Machine-Learning Partitioning (t-SNE and DBSCAN clustering) to analyze geohazard-environment linkages across Guangdong. Four principal conclusions emerge from this research:

• Spatial risk stratification: Guangdong exhibits distinct geohazard zones: high-risk mountainous northeast (Clusters 1–2), medium-risk transitional areas (Clusters 0, 3), and low-risk coastal plains (Cluster 4). This stratification provides a targeted management basis.
• Key controlling factors: Geohazard susceptibility is primarily governed by topographic steepness (LS factor), modulated by vegetation coverage (C_mean) and rainfall variability (R_slope). Soil conservation measures show location-dependent effectiveness, achieving optimal results when combined with vegetation restoration on moderate slopes.
• Climate change impacts: Increasing rainfall variability (R_slope) emerges as a significant emerging risk factor, particularly given climate projections for intensified precipitation extremes in South China.
• Management implications: High-risk zones: Prioritize engineering-slope stabilization combined with deep-rooted vegetation; Medium-risk zones: Implement vegetation-based solutions with continuous cover monitoring; All zones: Establish early-warning systems triggered by R_slope thresholds.