A machine learning-based study on ecological sensitivity and its driving factors in the karst landforms of Yangshuo County, Guilin

Introduction: Karst landscapes are characterized by fragile ecosystems due to their shallow soil layers, unique hydrological structures, and high habitat heterogeneity. The driving mechanisms of ecological processes within these systems are highly complex. Traditional evaluation methods (e.g., AHP, PCA), which are often based on linear assumptions, struggle to effectively capture complex mechanisms such as nonlinearity and high-order interactions within multi-factor interactions. This results in limited capacity for identifying and interpreting the driving factors of ecological sensitivity. Scientifically assessing this sensitivity is crucial for achieving regional sustainable development. This study takes Yangshuo County, Guilin --a typical karst area --as a case study.

Methods: It introduces the Self-Organizing Map (SOM), Random Forest (RF) model and the SHAP (SHapley Additive exPlanations) interpretability framework to evaluate ecological sensitivity based on a synthesis of 11 factors, including lithology, rocky desertification, and slope gradient.

Results: The results indicate that: (1) The ecological sensitivity in Yangshuo County can be classified into five distinct levels, predominantly dominated by vegetation-type sensitive areas and valley cultivated land sensitive areas. (2) The Random Forest model identified natural baseline factors, such as lithology, rocky desertification, and slope gradient, as the key drivers. (3) SHAP analysis further revealed non-linear interaction mechanisms among these factors. Crucially, it identified a "geological baseline - topographic dynamics - ecological process" cascading effect. This includes interactions such as steep slopes amplifying rocky desertification risks, the blocking effect of vegetation at the critical threshold of desertification, and the superposition and modification of natural factor influences by human activities.

Discussion: The "Self-Organizing Map-Random Forest-SHAP" (SOM-RF-SHAP) evaluation framework developed in this study provides a novel methodology for quantifying the complex driving mechanisms of ecological sensitivity in karst regions. The findings offer a scientific basis for ecological conservation and spatial planning in Yangshuo County and similar areas.

1 Introduction

Karst landscapes are surface and subsurface formations resulting from the long-term dissolution, erosion, and gravitational collapse of soluble rocks, such as limestone and dolomite, by groundwater and surface water. They are characterized by complex systems including rugged surface features like peak clusters, solution grooves, and sinkholes, as well as subterranean features such as caves and underground rivers (Chang et al., 2024). These landscapes record geological evolutionary history and serve as crucial archives for paleoclimate reconstruction and biodiversity hotspots. Their heterogeneous environments foster highly specialized ecosystems, providing vital evidence for global researchers studying climate change, hydrological cycles, and paleogeographical environments (Chen et al., 2024). As a critical component of global karst ecosystems, the heterogeneous environment and inherent fragility of karst landscapes have profound implications for regional ecological security and sustainable development (Dao-Xian, 2008).

Karst landscapes in China are primarily distributed across its western regions, including Guangxi, Guizhou, and eastern Yunnan, with a total area ranging from approximately 910,000 to 1,300,000 square kilometers. This makes it one of the largest karst regions in the world. These landscapes exhibit a latitudinal zonation across China, transitioning from tropical karst in the south, through subtropical karst, to temperate karst in the north. The western karst areas comprise arid region karst (in the northwest) and alpine frost karst (on the Qinghai-Tibet Plateau). Based on regional ecosystem service assessments and geo-ecological studies, the academic community widely recognizes karst environments as ecologically fragile zones, comparable in vulnerability to desert margins, and characterized by distinct sensitivity traits (D'ettorre et al., 2024). With accelerating urbanization, the conflict between human activities, such as tourism development, and the conservation of karst landscapes is intensifying. Consequently, the ecological security of karst areas has become a growing focus of concern for researchers and practitioners in related fields (Lu, 2024). Karst landscapes exhibit high ecological fragility due to their unique geological structures, such as high bedrock exposure rates and thin, discontinuous soil layers. Studies indicate that soil erosion rates in the karst areas of the Pearl River Basin can be 1.2 times higher than those predicted by conventional models, and the risk of rocky desertification is projected to intensify significantly under high carbon emission scenarios. As a typical karst peak cluster landscape, Yangshuo County exhibits a rocky desertification area comprising 18.7% of its total territory, with an average annual soil erosion modulus of 3800 t·ha⁻¹·yr⁻¹. These conditions impose significant constraints on its ecological carrying capacity. Hence, quantifying ecological sensitivity is crucial for ensuring regional ecological security.

In traditional ecological sensitivity assessments, linear models, notably the Analytic Hierarchy Process (AHP) and Principal Component Analysis (PCA), have been widely adopted. For instance, in a comprehensive evaluation of ecological sensitivity for a karst mountain city, the AHP method designated land use type (with a weight of 0.4) and vegetation coverage (with a weight of 0.4) as the dominant factors through subjective weighting. However, the weighting system in this approach is susceptible to expert cognitive bias (Dong et al., 2025). In an ecological security assessment of the karst rocky desertification area in Xiangxi, researchers selected 14 key influencing factors based on dynamic processes and extracted two principal components (P1 and P2) using PCA. The study demonstrated that while PCA effectively identified the spatial gradient patterns of rocky desertification drivers, it failed to elucidate the non-linear synergistic effects among these factors due to its inherent linearity assumptions (Liu and Li, 2012). The aforementioned limitations underscore the theoretical bottleneck of traditional methods in characterizing the complex driving mechanisms within heterogeneous karst surface systems. This study introduces the Random Forest algorithm, whose advantages include: the capability to process high-dimensional data (e.g., variables in Yangshuo County involving 12 nationally protected plant species and 5 soil types), automatic screening of key driving factors, and quantification of feature importance. When integrated with the SHAP method, it can further unravel the interaction mechanisms among factors. For instance, in a study of the Chang-Zhu-Tan urban agglomeration, the Random Forest model revealed that the interaction term between “land use intensity and slope” had an explanatory power 2.3 times greater than that of any single factor (Chen Y. et al., 2025). Furthermore, SHAP value visualization demonstrated that in areas of high-intensity development, the ecological risk remained significantly higher than in steep slope areas with low development intensity, even when the terrain was gentle. This “algorithm + interpretability” framework transcends the linear assumptions of traditional methods, providing a more refined tool for ecological sensitivity assessment and simultaneously offering an ideal scenario for validating the generalization capability of machine learning models (Jiang, 2024; Ji et al., 2023).

The “SOM-RF-SHAP” evaluation framework developed in this study introduces a core innovation by establishing a complete closed-loop analytical pathway from “pattern recognition” to “mechanistic interpretation.” In contrast to studies that rely solely on clustering or classification models, our framework employs the Self-Organizing Map (SOM) for unsupervised clustering, overcoming the dependence on predefined parameters inherent in traditional threshold-based methods or algorithms such as K-Means, thereby enabling a more objective identification of the baseline pattern of ecological sensitivity from high-dimensional data. The SOM clustering results are subsequently used as target variables to drive the Random Forest model, effectively quantifying factor importance and circumventing the inherent biases of subjective weighting methods like AHP. Building on this, the SHAP interpretability model is incorporated to penetrate the “black box” of the Random Forest, not only validating the global importance ranking of factors but, more critically, revealing the nonlinear operational mechanisms and interaction effects of different factors under specific contexts. This integrated framework systematically addresses three core questions in ecological sensitivity research in karst areas: “how patterns differentiate,” “which factors dominate,” and “how factors interact,” thereby providing new analytical depth and methodological support for understanding its formation mechanisms.

2 Materials and methods

2.1 Overview of the study area

Yangshuo County is situated in the northeastern part of the Guangxi Zhuang Autonomous Region, China, and falls under the administration of the internationally renowned tourist city of Guilin. It lies between 110°13′–110°40′E longitude and 24°38′–25°04′N latitude. The county administers 6 towns and 3 townships, covering a total administrative area of 1,436 square kilometers. The terrain is dominated by typical karst peak forest and peak cluster landforms, with higher elevations in the northwest and lower in the southeast. The Li River traverses the entire region, forming a unique landscape known as the “Hundred-Mile Gallery.” Characterized by a mid-subtropical monsoon climate, the county experiences a mean annual temperature of 19 °C, receives an average annual precipitation of 1,596 mm, and has a frost-free period of 308 days. It boasts a forest coverage rate of 75.8% and harbors over 1,000 species of vascular plants, including 12 nationally protected species. The predominant soil types are red soil and limestone soil, which are rich in organic matter (Figure 1).

Figure 1

Figure 1. Geographical location of the study area, Yangshuo County.

As a world-renowned tourist destination, tourism serves as the pillar industry of Yangshuo County. However, constrained by its karst terrain, the region exhibits inherent ecological fragility, with rocky desertification affecting 18.7% of its land area and an average annual soil erosion modulus reaching 3800 t·km⁻². The intense tourism development conflicts sharply with this fragile ecological environment, making Yangshuo a typical area for studying the interaction between ecological sensitivity and human activities—a veritable natural laboratory for karst ecological sensitivity. The county possesses a complete geomorphological sequence, serving as a representative specimen of karst evolution. Furthermore, its ecosystem endures dual pressures from both rocky desertification and high-intensity tourism development, constituting a compound case study of human-land interaction (Zhang, 2023). Coupled with its role as a core area for innovative governance practices in Li River ecological restoration, these attributes collectively establish the regional typicality and representativeness underpinning our case selection (Tu et al., 2024).

2.2 Research methodology

This study established an integrated assessment framework combining a Self-Organizing Map, Random Forest, and the SHAP interpretability model (Figure 2). The framework begins with the preprocessing of multi-source geospatial data, followed by unsupervised identification of spatial patterns in ecological sensitivity via SOM clustering. The resultant clusters are then used as target labels to train a Random Forest model, which quantifies the global importance of each driving factor. Finally, the SHAP model is employed to deconstruct the Random Forest predictions, revealing the non-linear effects and interaction mechanisms of the driving factors at both local and global levels, thereby completing an analytical chain from pattern discovery to mechanistic interpretation.

Figure 2

Figure 2. Research technical framework flowchart.

2.2.1 Data sources and preprocessing

Digital Elevation Model (DEM) data with a 30-m resolution was sourced from the Geospatial Data Cloud (http://www.gscloud.cn). Slope and aspect data were de-rived from this DEM using the “Slope/Aspect Tool” in ArcGIS 10.6. Runoff data was generated through hydrological analysis of the elevation data, and Euclidean Distance calculation was applied to create a raster layer representing distance to runoff. River vector data was obtained from Tianditu (https://www.tianditu.gov.cn), and similarly, Euclidean Distance calculation was used to generate a raster layer of distance to rivers. Land use data was acquired from the CLCD database of Wuhan University (http://www.geososo.com). Vegetation cover type data was provided by the National Cryosphere Desert Data Center (http://www.ncdc.ac.cn). Fractional Vegetation Cover (FVC) data was processed on the Google Earth Engine (GEE) platform. Lithology data was obtained from the Earth Resource Data and Analysis Cloud Platform (http://www.geodata.cn). Soil erosion and rocky desertification data were sourced from the Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences (https://www.ecosystem.csdb.cn). The vector boundary of the study area was provided by the Resource and Environment Science and Data Center of the Chinese Academy of Sciences (http://www.resdc.cn).

Using ArcGIS 10.8 software, the coordinate systems of all factor raster datasets were unified to WGS_1984_UTM_Zone_49N and spatially resampled to a consistent 500 m × 500 m resolution. All factor rasters were subsequently normalized to a value range between 0 and 1. A fishnet was constructed and converted to point vector for-mat, and the attribute information from each factor raster layer was extracted to these corresponding points. Following the removal of data points with incomplete information, a total of 5,709 valid sample points were obtained for the subsequent analysis.

2.2.2 Selection of influencing factors

Existing studies on karst ecological sensitivity predominantly focus on natural elements (e.g., slope gradient, lithology, degree of rocky desertification), while the integration of socio-economic factors remains relatively limited (Dong et al., 2025). Given that the county scale plays a pivotal bridging role in the ecological civilization construction system, where the state of the ecological environment profoundly constrains regional sustainable development capacity, making it a critical arena for coordinating socio-economic development and ecological conservation, therefore, ecological sensitivity research targeting county-level units necessitates a foundation in their specific regional characteristics and the selection of key influencing factors that are both dominant and demonstrate strong regional applicability (Wang and Shang, 2025; Yang et al., 2025).

This study adhered to the principles of systematicity, regional adaptability, and data accessibility, while simultaneously ensuring the processed data met the input requirements for machine learning models. A bibliometric analysis was conducted to systematically review domestic karst ecological sensitivity studies from the past 5 years, extracting high-frequency factors and consolidating similar categories (Tu et al., 2024; Chen L. et al., 2025; Yuan, 2025; Qiu et al., 2024; Li et al., 2024; Wang et al., 2024; Lin, 2024; Liu, 2024; Wei, 2024; Chen, 2024; Jia et al., 2024; Zhu et al., 2023; Fan, 2023; Gong, 2023; LIANG et al., 2023; Tian and Wang, 2025; YING et al., 2023). Building upon this foundation and closely aligning with the natural geographical context and human activity intensity of Yangshuo County, Guilin, as a typical karst region, a final set of 11 influencing factors for ecological sensitivity was selected: lithology, rocky desertification index, slope gradient, aspect, elevation, distance to rivers, distance to runoff, fractional vegetation cover (FVC), vegetation type, land use type, and soil erosion intensity. These factors comprehensively represent six major categories: geology, topography, hydrological processes, vegetation status, human activities, and soil erosion (Table 1), collectively characterizing the combined driving forces of the natural baseline and anthropogenic disturbances on the fragile karst ecosystem. The spatial distribution of each influencing factor is shown in (Figure 3).

Table 1

Table 1. Influencing factors for ecological sensitivity in the karst landscape of Yangshuo county, Guilin.

Figure 3

Figure 3. Spatial Distribution of Ecological Sensitivity Influencing Factors in Yangshuo County, Guilin. (a) Lithology; (b) Degree of rocky desertification; (c) Slope sensitivity; (d) Aspect sensitivity; (e) Elevation sensitivity; (f) Water distance; (g) Runoff distance; (h) Vegetation coverage; (i) Vegetation type; (j) Land use; (k) Sensitivity to soil erosion.

2.2.2.1 Lithology

Yangshuo County exhibits a diverse range of exposed lithologies, including shale, sandstone, dolomitic limestone, and gravel/clay layers, among others. The most extensive are the gravel/clay layers, while the gravel/sand/clay layers cover the smallest area. Carbonate rocks, such as limestone and dolomitic limestone, constitute a significant proportion, indicating a well-developed karst landscape with prominent interbedding phenomena, characterized by a complex alternation of clastic and carbonate rocks. This complex lithological assemblage influences groundwater permeability, dissolution rates, and surface morphology. Areas underlain by limestone and dolomite are prone to dissolution, collapse, and groundwater pollution, exhibiting weak ecological resto-ration capacity, making them core zones for ecological conservation and rocky desertification control. Carbonate rock areas containing siliceous or chert bands remain quintessential karst environments. Interbedded mudstone and sandstone areas are susceptible to localized soil erosion or landslides due to lithological contrasts. Intercalated insoluble layers (e.g., chert bands, mudstone layers) act as relative aquitards, influencing the flow direction of underground rivers and the location of aquifers, often leading to spring emergence which forms critical ecological water sources. Shale and sandstone areas generally possess thicker soil layers, better water retention capacity, and relatively higher ecological stability (Figures 2a).

2.2.2.2 Rocky desertification index

The area affected by rocky desertification in the middle and lower reaches of the Li River is significantly larger than in the upper reaches. Yangdi Township exhibits the most extensive rocky desertification, with Fuli Town, Yangshuo Town, Xingping Town, and Baisha Town also identified as key distribution areas. The widespread distribution of carbonate rocks in Yangshuo County provides the material basis for rocky desertification formation. These highly soluble rocks, under the effects of rainfall dissolution and water erosion, lead to the rapid loss of topsoil, resulting in bedrock exposure. Although natural factors form the foundation, unsustainable engineering activities such as slope cutting for construction, road excavation, and certain agricultural practices exacerbate soil erosion, trigger geological hazards like collapses, and consequently accelerate the process of rocky desertification (Figures 2b).

2.2.2.3 Slope gradient

The maximum slope gradient in Yangshuo County reaches 75.8667°. Gentle and flat slopes (<15°) are primarily distributed along the banks of major rivers such as the Li River and the Yulong River, as well as in the solution plains and depressions be-tween hills. These areas concentrate most towns, villages, and farmland. Slopes and steep slopes (15°–35°) are mainly found on hill-foot slopes, hills, and the flanks of continuous peak clusters, forming a transitional zone from gentle plains to steep ridges. This terrain is typically covered by shrubland or secondary forests. Very steep and precipitous slopes (>35°) constitute the main body of the karst peak forest landscape, characterized by extensive rock exposure. Vegetation here consists predominantly of lithophytic and drought-tolerant shrubs, grasses, or scattered trees, distributed on the cliffs and steep slopes of karst peaks along the river basins (Figures 2c).

2.2.2.4 Aspect

Macroscopically, due to the conical shapes of the karst peak clusters where slopes face all directions, the distribution of sunny (south-facing), semi-sunny (southeast, southwest), shady (north-facing), and semi-shady (northeast, northwest) slopes is generally balanced, with no single aspect being overwhelmingly dominant. Different aspects on individual peaks create contrasting microenvironments with significant microclimatic differences. For instance, steep south-facing slopes receive substantially different solar radiation compared to gentle south-facing ones. In regions like Yangshuo, characterized by extensive steep and precipitous slopes, the microclimatic effects of aspect are amplified (Figures 2d).

2.2.2.5 Elevation

The karst landscape of Yangshuo, shaped by prolonged dissolution and denudation, exhibits distinct topographic tiers, with the highest point at 1,686 m and the lowest at −15 m. Low-elevation plains/valleys (mostly below 150 m) consist of alluvial plains along rivers like the Li and Yulong and large solution depressions, featuring flat and open terrain. Karst peak forests typically range between 150 and 400 m in elevation, comprising relatively isolated or clustered limestone peaks rising abruptly from the plains, forming the iconic “peak forest plain” landscape. Continuous karst peak clusters ascend to 400–600 m or higher; these interconnected mountains with rugged summits present more complex topography and are often inaccessible. Notably, flat valleys and steep peaks lie in immediate proximity within Yangshuo’s karst terrain, creating dramatic local relief often exceeding 200–300 m. This significant elevation difference is a primary driver causing sharp variations in ecological factors (e.g., temperature, humidity) over very short distances (Figures 2e).

2.2.2.6 River buffer zone

The distribution of the river buffer zone primarily manifests as a dendritic net-work extending along trunk streams like the Li River and Yulong River and their tributaries. The buffer width varies with stream order and topographic relief, being generally wider in plain and basin sections and narrower in gorge segments traversing peak clusters. These buffer zones form green corridors connecting different ecological patches and serve as vital pathways for water-soil exchange, species migration, and energy flow. The conservation status of these river buffers is directly linked to water environment security and biodiversity maintenance in the karst areas of Yangshuo County, making them critical targets for ecological sensitivity assessment and spatial regulation (Figures 2f).

2.2.2.7 Runoff buffer zone

The runoff buffer zone in Yangshuo County extends beyond perennial rivers, encompassing all water movement pathways, including seasonal gullies, foothill diffuse flow areas, and surface convergence zones above karst underground river systems. Its spatial form is not distinctly linear but rather appears as patches and networks con-trolled by micro-topography, often discontinuous. In karst regions, significant runoff is diverted underground through features like fissures and sinkholes, rendering the location and extent of these surface buffers more concealed and dynamic, and not entirely overlapping with the surface river network (Figures 2g).

2.2.2.8 Fractional vegetation cover

The spatial distribution of Fractional Vegetation Cover (FVC) in Yangshuo County exhibits a strong correlation with the karst geomorphic pattern. Overall, areas with higher FVC are primarily concentrated on the tops and shady slopes of continuous peak clusters with minimal human disturbance, as well as some mountainous areas undergoing ecological restoration. Conversely, lower FVC areas are scattered around towns, villages, and along major transportation routes, with particularly sparse vegetation on sunny slopes and steep hillsides where limestone exposure is severe. This distribution pattern directly reflects the combined effects of natural constraints and anthropogenic disturbance (Figures 2h).

2.2.2.9 Vegetation type

The distribution of vegetation types in Yangshuo County exhibits distinct zonation patterns and clear signatures of human disturbance. Broad-leaved forests are primarily distributed on shady slopes or hilly areas with minimal human disturbance and favorable hydrothermal conditions, representing well-preserved ecosystems within peak-cluster depressions. Coniferous forests are often found on relatively infer-tile sunny slopes or the mid-upper sections of hills, frequently as plantations or secondary forests. Tussock, primarily as a secondary type following forest destruction, is widely distributed in low-elevation plains and areas with frequent human activities. Bush fallow is scattered throughout farming regions, representing an early stage of vegetation succession (Figures 2i).

2.2.2.10 Land use type

In Yangshuo County’s land use pattern, Forest—representing the largest proportion—is predominantly distributed in the northeastern and southwestern regions. Cropland is extensively found in the relatively flat valleys and plains along rivers such as the Li River and Yulong River, as well as on gentle foothill slopes, constituting concentrated areas for agricultural production and village distribution. Shrubland and Grassland are interspersed between forest and cropland areas, widely scattered across mountainous regions. Water bodies primarily include the Li River, Yulong River, and other river systems along with their peripheral zones, whose banks form critical eco-logical buffer areas. Built-up land is mainly concentrated in Yangshuo County’s urban core, various tourist towns, rural settlements, and along major transportation corridors (Figures 2j).

2.2.2.11 Soil erosion intensity

The spatial distribution of soil erosion intensity in Yangshuo County exhibits a high degree of coupling with the geological background and human activities. Moder-ate-to-severe soil erosion is primarily concentrated in the central and southwestern regions. These areas often correspond to outcrops of weakly erosion-resistant clastic rocks, such as clay layers and sandy shale, or to hill-foot slopes within limestone ter-rains that have relatively thicker soils but experience intensive cultivation. The erosion patterns frequently manifest in patchy or linear forms, showing significant spatial overlap with areas disturbed by engineering activities like steep slope cultivation, or-chard development, and infrastructure construction. In contrast, the tops of typical peak clusters, characterized by extensive bedrock exposure, actually exhibit lower soil erosion rates (Figures 2k).

2.2.3 Self-Organizing Map (SOM) model

This study employed the Self-Organizing Map (SOM) neural network model to perform an unsupervised clustering analysis of ecological sensitivity in Yangshuo County. As an efficient unsupervised clustering algorithm, the SOM’s advantage lies in its ability to project and visualize high-dimensional, nonlinear input data (specifically, the 11 ecological sensitivity influencing factors in this study) onto a low-dimensional space (typically two-dimensional) in a topology-preserving manner (Kohonen, 2013). Geographical units possessing similar characteristics in the original high-dimensional space are consequently assigned to the same or adjacent clusters in the SOM output. This approach, which does not rely on predefined class labels, objectively reveals the inherent grouping patterns of ecological sensitivity types within Yangshuo County based purely on the data’s intrinsic structure, thereby avoiding the subjectivity associated with manually defined thresholds.

To assign clear ecological meanings to the numerical results generated by the unsupervised SOM clustering, this study further adopted a systematic, data-driven interpretation process for naming the cluster categories, ensuring the objectivity and ecological interpretability of the labels. The process began by spatially visualizing the SOM clustering results in the ArcGIS platform and overlaying them with high-resolution remote sensing imagery and the spatial distribution maps of the 11 impact factors (Figure 3). This allowed for the precise identification of the actual land use types (such as forest land, cropland, and construction land) and the primary geomorphological units (such as peak-cluster depressions, river valleys, and plains) dominated by each numerical category. Based on this spatial analysis, the one or two most distinctive dominant ecological attributes of each category were extracted and combined with the “sensitivity” characteristic to form the category names. For example, the “Alpine Forest Land Sensitive Category” refers to categories predominantly distributed in high-altitude forest areas, where sensitivity stems from topographical and vegetation attributes, while the “Population Aggregation Sensitive Category” denotes areas characterized by high-intensity human settlement activities as the dominant attribute. This methodology ensures that the transformation from numerical clusters to ecological concepts is based on spatial evidence and data-driven objective inference.

2.2.4 Random Forest algorithm

In order to objectively quantify the relative contribution of each influencing factor to the spatial heterogeneity of ecological sensitivity in Yangshuo County and to overcome the limitations inherent in subjective weighting methods like the Analytic Hierarchy Process (AHP), this study employed the Random Forest (RF) algorithm to evaluate factor importance (Breiman, 2001). Random Forest is an ensemble machine learning algorithm composed of multiple decision trees. It possesses powerful capabilities for handling high-dimensional data, resisting overfitting, and outputting feature importance rankings. The algorithm utilizes Bootstrap sampling to generate a training set for each tree and randomly selects a subset of features for node splitting. This “dual randomness” enables the model to robustly evaluate the average contribution of each factor in predicting ecological sensitivity categories. It employs the decrease in the Gini Index or the reduction in Out-Of-Bag (OOB) Error to quantify feature importance. In this study, a higher importance value for a specific factor indicates its stronger discriminative power in distinguishing between different ecological sensitivity classes.

This study used the results from the SOM clustering in Section 2.2.3 (namely, the ecological sensitivity class label for each sample point) as the prediction target for the Random Forest model. The entire sample dataset was randomly split into a training set and a test set with a ratio of 70%–30%. The Random Forest Classifier from the Scikit-learn library on the Python platform was utilized to build the model. The model was trained using the training set, and its performance was validated using metrics such as test set accuracy, precision, and AUC value. Provided that the model performance met the required standards, the importance ranking of the 11 influencing factors was output, thereby enabling the scientific identification of the dominant and secondary factors driving the spatial differentiation of ecological sensitivity in Yangshuo County.

2.2.5 SHAP for quantifying factor contributions

While the Random Forest algorithm provides a global importance ranking of factors, it does not elucidate how individual factors influence the ecological sensitivity classification of specific regions, nor does it reveal potential interaction effects between factors. To address this limitation, this study introduces the SHAP (SHapley Additive exPlanations) method for model interpretability analysis (Lundberg and Lee, 2017). The SHAP method, rooted in the concept of Shapley values from cooperative game theory, equitably distributes the contribution value of each feature to individual prediction outcomes (Ji et al., 2023). This enables clear and interpretable explanations at both global (overall model behavior) and local (individual instance prediction) levels. In the ecological sensitivity analysis of Yangshuo County, the SHAP method was introduced with the dual objectives of: 1) providing an in-depth interpretation of the machine learning model’s prediction mechanisms, specifically quantifying the direction and magnitude of the effect that changes in individual driving factors have on ecological sensitivity classification; and 2) detecting and quantifying the interaction effects between different driving factors (Zhang et al., 2024; Eker and Aydın, 2024). For instance, it aims to reveal whether a non-linear synergistic effect exists between “rocky desertification” and “vegetation coverage” – specifically, whether the co-occurrence of “high rocky desertification” and “low vegetation coverage” leads to a sharp increase in ecological sensitivity.

By analyzing the SHAP values across all samples, this study not only validates the global feature importance derived from the Random Forest model but also provides deeper insights into the positive or negative influences of each factor on the five ecological sensitivity classes (e.g., alpine woodland sensitive areas, valley cultivated land sensitive areas) and their complex, non-linear dependencies. Thereby, it offers a mechanistic explanation for formulating differentiated and targeted strategies for ecological conservation and spatial planning.

3 Results

3.1 Analysis of ecological sensitivity clustering results

To determine the initial candidate range for cluster analysis, the K-Means algorithm was first employed, and its elbow curve was plotted (Figure 4). As shown, the within-cluster sum of squares (Inertia) decreases sharply as the number of clusters K increases from 1 to 5; beyond K > 5, the rate of decrease in inertia slows down significantly, forming an elbow-like bend. According to the elbow criterion, the K value at this bend (K = 5) often suggests the optimal number of clusters. To facilitate a comprehensive comparison, K = 4 and K = 6, adjacent to this elbow point, were also included alongside K = 5 for subsequent rigorous quantitative evaluation.

Figure 4

Figure 4. Preliminary Selection of Cluster Numbers using the K-Means Elbow Method.

Following the preliminary identification of the candidate range (K = 4, 5, 6),to determine the optimal number of clusters, this study compared scenarios with cluster numbers (K) set to 4, 5, and 6 using both Self-Organizing Map (SOM) and K-Means algorithms. A quantitative assessment was conducted based on three internal validation metrics: the Silhouette Coefficient, Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CH) (Table 2). The results indicated that the SOM algorithm consistently and significantly outperformed the K-Means algorithm across all metrics. Among the SOM results, the configuration with K = 5 achieved the best Davies-Bouldin Index (2.2122), demonstrating an optimal balance between intra-cluster compactness and inter-cluster separation. Although its Silhouette Coefficient (0.1238) and CH Index (3724.109) were slightly lower than those for K = 4, they remained at high levels, indicating good clustering structure consistency and inter-cluster variance explanation. Comprehensive consideration concluded that K = 5 achieved the optimal balance between the statistical validity of the cluster structure and practical spatial interpretability, thus being determined as the final number of clusters.

Table 2

Table 2. Comparison of clustering validation metrics.

To visually verify the structure of the clustering results in the feature space, a scatter plot based on the first two principal components was generated, with clusters for K = 5 color-coded (Figure 5). As shown, the five categories exhibit a clear separation trend in the PCA space dimensionally reduced from the original 11 ecological factors. The centers of the categories are distinctly distributed with limited overlapping areas. This visualization result confirms from the origin of the data structure that the five ecological sensitivity categories obtained by SOM clustering are not random but stem from intrinsic differentiations within the high-dimensional feature space of the driving factors, thereby visually reinforcing the rationality of selecting the five-cluster solution.

Figure 5

Figure 5. Visualization of SOM Clustering Results for Ecological Sensitivity based on Principal Component Analysis.

Subsequently, the ecological sensitivity category labels generated from this SOM clustering result (K = 5) were used as the prediction target for the Random Forest model for in-depth driving factor analysis. Furthermore, following the principles of conjugacy and integrity, fragmented patches in the initial clustering output were processed using ArcGIS 10.8 software, ultimately yielding the refined SOM network clustering results (Figure 6) and the distribution of ecological sensitivity categories across townships (Table 3).

Figure 6

Figure 6. Distribution of ecological sensitivity classes by township in Yangshuo county, Guilin.

Table 3

Table 3. Distribution of ecological sensitivity classes by township.

The ecological sensitivity in Yangshuo County, Guilin, is predominantly characterized by Class III and Class IV areas (Table 4). Class III areas, covering the largest portion at 25.24% of the total county area, are primarily distributed in Jinbao Town-ship and Gaotian Town in the southwest, Puyi Township in the south, and Xingping Town in the north, and Fuli Town in the east. Class IV areas account for 24.61% of the total area, mainly located in Baisha Town in the central part, Putao Town in the northwest, and also Fuli Town in the east. Class V areas constitute 20.74% of the county’s area and are primarily found in Yangshuo Town in the central region, as well as Yangdi Township and Xingping Town in the north. The primary distribution characteristics shared by Class IV and V areas include their location in low-elevation river valleys and plains, covering the main cultivated land areas of Yangshuo County. These areas exhibit moderate to high levels of rocky desertification, relatively developed economies, and higher population densities. In contrast, Class II and Class I areas ac-count for the smallest proportions, at 12.13% and 17.28% of the total area, respectively. They are distributed in southwestern Jinbao Township and the northern part of Xing-ping Town, typically in high-elevation regions. These areas feature low levels of rocky desertification, land use dominated by forest, high vegetation coverage, abundant for-est resources, and low population density.

Table 4

Table 4. Area and proportion of the five ecological sensitivity classes in Yangshuo county.

3.2 Analysis of overall feature importance

This study performed hyperparameter optimization for the Random Forest model using Grid Search and Randomized Search, with a focus on the impact of the number of decision trees($n_{\text{estimators}}$)and the maximum depth ($\max \_\text{depth}$) on model performance. The macro-F1 score was set as the primary optimization target because it serves as a comprehensive evaluation metric for classification models, particularly in multi-class or imbalanced classification tasks, more effectively reflecting model robustness and predictive capability across all classes than accuracy alone. The analysis results (Table 5) showed that the model achieved the highest macro-F1 score of 0.9310, along with an accuracy of 0.9280, when the number of estimators was set to 500 and the maximum depth was set to 20. This optimal combination indicates that with an ensemble size of 500 trees, the model’s variance is effectively reduced, avoiding overfitting, while limiting the maximum depth to 20 layers ensures the tree’s expressive power (Bias) and effectively prevents individual decision trees from over-learning noise. Therefore, the parameter configuration of $n_{\text{estimators}}=500$ and $\max \_\text{depth}=20$ was determined as the optimal hyperparameter combination based on empirical data and the principle of performance metric maximization, providing a reliable academic basis for the deployment of the final model.

Table 5

Table 5. Performance Comparison of Random Forest Hyperparameter Optimization.

The performance of the final model, constructed based on this optimal parameter combination, was evaluated using the confusion matrix (Figure 7) and the ROC curve (Figure 8). The model achieved a test set accuracy of 0.928. As shown in Figure 8, the ROC curves plotted using the One-vs-Rest strategy reveal that the Area Under the Curve (AUC) for all five ecological sensitivity categories exceeded 0.984, with two categories achieving a perfect 1.000, and the model’s macro-average AUC reached 0.998. This robustly demonstrates the exceptional classification performance and powerful discriminatory capability of the optimized model, establishing a solid foundation for its subsequent use in reliable feature importance calculation.

Figure 7

Figure 7. Random forest confusion matrix.

Figure 8

Figure 8. Random forest ROC curve.

As shown in Figures 9, 10, six features - lithology, rocky desertification, river buffer, slope gradient, soil erosion, and vegetation type - all demonstrated importance scores exceeding 10%. Among these, lithology approached 20%, identifying it as the most influential factor. In contrast, land use importance was below 1%, representing the least significant factor. The importance scores for aspect, elevation, runoff buffer, and vegetation coverage were all approximately ≤2.5%. Based on the scatter plot of importance values, the 11 features can be categorized into two distinct groups: the first six high-importance features form Category 1, playing a dominant role in ecological sensitivity clustering, while the remaining five low-importance features constitute Category 2.

Figure 9

Figure 9. Scatter plot of feature importance from random forest for ecological sensitivity factors in Yangshuo county.

Figure 10

Figure 10. Bar plot of feature importance from random forest for ecological sensitivity factors in Yangshuo county.

This analysis reveals that ecological sensitivity in Yangshuo County is closely associated with natural geological factors. The six Category 1 features share the commonality of representing inherent surface attributes and natural resistance to disturbance. These features characterize the ecosystem’s “baseline” vulnerability or stability under natural conditions, emphasizing sensitivity to geological, topographic, hydrological sensitive zones, and natural vegetation coverage against erosion processes. The five Category 2 features can be summarized as representing spatial patterns, environmental gradients, and anthropogenic influence intensity, describing the spatial distribution and variation patterns of ecological factors and their resulting micro-environmental differences. These features primarily reveal spatial heterogeneity, gradient changes in environmental conditions, and the degree of pressure and impact from human activities. They determine the spatial distribution pattern of ecological sensitivity and its actual manifestation following anthropogenic disturbance.

The disparity in importance between these two categories of ecological sensitivity clustering features reflects their close relationship with the unique geological, hydrological, and ecological characteristics of the karst system, consistent with Yangshuo’s status as a typical karst landscape. The characteristic peak cluster-depression landscape in Yangshuo exhibits a dual surface-underground structure, resulting in surface water scarcity and vegetation dependence on soil moisture in rock fissures, making lithology and vegetation type critical bottlenecks for ecological restoration. When slope exceeds 15°, soil loss rates increase exponentially. The higher slope gradients in the middle and lower reaches of the Li River in Yangshuo create a positive feedback loop: slope → erosion → rocky desertification. Research indicates that when the rocky desertification rate exceeds 20%, ecosystem services decline precipitously. Consequently, the early warning value of Category 1 features, particularly lithology, slope, and rocky desertification, becomes particularly significant (Zhu, 2021). The Category 1 features directly characterize the core mechanisms of rocky desertification: lithology (impedes soil formation), slope gradient (facilitates soil loss), and vegetation type (determines soil stabilization capacity). Therefore, they emerge as the dominant factors governing sensitivity. In contrast, the effects of Category 2 features are constrained by the geological context of the karst system. For instance, the influence of aspect on moisture distribution is diminished by substantial subsurface leakage, while the ecological impact of land use intensity is buffered by the inherent resistance of the lithology.

3.3 SHAP summary analysis

The SHAP summary plot for the ecological sensitivity influencing factors in Yangshuo County (Figure 11) illustrates the contribution of each factor towards classifying sample points into specific sensitivity categories. A larger absolute SHAP value indicates a greater contribution to the clustering outcome. A positive SHAP value signifies that the specific feature value has a positive influence on the classification, while a negative value indicates a negative influence.

Figure 11

Figure 11. SHAP summary analysis plots for ecological sensitivity influencing factors in Yangshuo County: (a) Sensitive category of alpine forest land (Class I); (b) Slope sensitive category (Class II); (c) Vegetation type sensitive (Class III); (d) Sensitive category of cultivated land in river valleys (Class IV); (e) Population aggregation sensitive category (Class V).

As shown in Figure 8a, lithology is the most contributing factor for this category (Class I), followed by the degree of rocky desertification and soil erosion intensity. Lower feature values of lithology correspond to higher SHAP values, demonstrating a stronger positive contribution towards classifying a sample into this category. Conversely, higher feature values of soil erosion intensity and rocky desertification correspond to higher SHAP values, also indicating a stronger positive contribution towards this classification. The spatial distribution of this type shows high overlap with the alpine forest land areas in Yangshuo County; therefore, it is designated as the Sensitive Category of Alpine Forest Land (Class I).

As shown in Figure 8b, rocky desertification is the most contributing factor for this category (Class II), with slope gradient and lithology ranking as the second and third most important factors. In contrast to the previous category, lower feature values of rocky desertification correspond to higher SHAP values, indicating a stronger positive contribution toward classifying a sample into this class. Conversely, higher feature values of slope gradient correspond to higher SHAP values, demonstrating a positive contribution to this classification. The approximate ranking of contribution from the remaining factors is highly similar to the previous category. The most distinguishing factors between these two categories are slope gradient and aspect. Furthermore, comparison with the spatial distribution map of ecological sensitivity influencing factors (Figure 2) reveals that the distribution of this class shows high spatial overlap with areas at high risk of soil erosion. Based on the above analysis, this class is designated as the Slope Sensitive Category (Class II).

As shown in Figure 8c, rocky desertification is the most contributing factor for this category (Class III). However, unlike the previous class, higher feature values of rocky desertification correspond to higher SHAP values here. Rocky desertification, slope gradient, vegetation type, lithology, and soil erosion intensity are the top five contributing factors, in that order. For all these factors, higher feature values correspond to higher SHAP values, indicating a stronger positive contribution toward classifying a sample into this class. The key characteristic distinguishing this class from the other four is that it encompasses the greatest diversity of vegetation types. Therefore, this class is termed the Vegetation Type Sensitive Category (Class III).

As shown in Figure 8d, the river buffer zone is the most contributing factor for this category (Class IV). Higher feature values of the river buffer correspond to higher SHAP values, indicating a positive influence on this classification. Vegetation type and soil erosion intensity are the second and third most important factors. For these two factors, however, lower feature values correspond to higher SHAP values, showing a positive contribution. A key distinction between this class and the other four is that the land use type factor is not among the lowest in terms of contribution here. Further spatial comparison reveals that the distribution of this class highly coincides with cultivated land located in the river valleys of Yangshuo County. Consequently, this class is designated as the Sensitive Category of Cultivated Land in River Valleys (Class IV).

As shown in Figure 8e, slope gradient is the most contributing factor for this category (Class V). Lower feature values of slope correspond to higher SHAP values, demonstrating a positive influence on classification into this class. The river buffer zone and soil erosion intensity are the second and third most important factors. For these factors, lower feature values also correspond to higher SHAP values, indicating a positive contribution. The distribution characteristics of this class include locations in valley plains with low slope gradients and low elevations, situated within the highly concentrated population centers of various townships. Therefore, this class is named the Population Aggregation Sensitive Category (Class V).

3.4 SHAP interaction analysis

For Class I - Alpine Forest Land Sensitive Category: When the rocky desertification value is low, the absolute SHAP values of the slope factor are generally lower than when the rocky desertification value is high. This indicates that higher rocky desertification values enhance the contribution of the slope factor (Figures 12a). When elevation values are low, the feature values of vegetation coverage are relatively high, and the SHAP values for vegetation coverage show considerable variation, ranging from 0.1 to −0.3. In contrast, when elevation values are high, regardless of changes in vegetation coverage, the SHAP values for vegetation coverage remain between −0.1 and 0. Overall, the variation in SHAP values for these factors is smaller compared to others. (Figures 12b). Based on previous GIS analysis (Figure 3) and cluster analysis (Figure 6), the distribution areas of this category are characterized by high-elevation forest land with high vegetation coverage and altitude. Therefore, the interaction characteristics between elevation and vegetation coverage align with these findings: the overall absolute SHAP values are small, indicating a minor contribution to the Class I outcome, which is consistent with the features of this alpine forest land sensitive category.

Figure 12

Figure 12. SHAP feature dependency plots for ecological sensitivity influencing factors in alpine forest land sensitive category (Class I): (a) Interaction dependency of SHAP values between slope and rocky desertification; (b) Interaction dependency of SHAP values between vegetation coverage and elevation.

For Class II - Slope Sensitive Category, previous analysis identified rocky desertification, slope gradient, and lithology as the three most significant contributing factors, with soil erosion ranking fourth. These four factors exhibit a tightly coupled, logically progressive causal relationship. Lithology serves as the fundamental controller, particularly soluble bedrock (e.g., carbonate rocks like limestone and dolomite), forming the starting point and material basis of the entire causal chain. These rocks are highly susceptible to hydro-dissolution, developing karst features such as fissures, sinkholes, and subterranean rivers. Concurrently, the pedogenesis rate from soluble bedrock is exceptionally slow (requiring millennia to form merely 1 cm of soil), resulting in inherently thin, impoverished, and discontinuous soil cover.

Lithology, in turn, dictates slope gradient. Dissolution of soluble rocks creates highly rugged topography including peak clusters, depressions, and gorges, producing naturally steep slopes. Slope gradient acts as the primary dynamic driver of soil erosion. Steep slopes exhibit poor surface stability and pronounced gravitational effects, providing substantial potential energy for soil detachment and transport. Increasing slope angles accelerate surface runoff velocity during rainfall events, enhancing its soil scouring capacity. Steep gradients also reduce water residence time on surfaces, di-minishing infiltration and increasing overland flow that carries away topsoil. Furthermore, soils on steep slopes are inherently vulnerable to mass wasting events like landslides and collapses. Consequently, this category’s distribution shows pronounced spatial coincidence with high soil erosion risk zones.

Soil erosion constitutes the direct process culminating in rocky desertification. In-tensive erosion strips the thin soil mantle and weathered debris from bedrock surfaces. When soil loss rates drastically outpace bedrock weathering rates, extensive bedrock exposure occurs. This degenerative process—characterized by soil depletion, bedrock emergence, and severe degradation of land productivity—defines rocky desertification. Ultimately, rocky desertified landscapes lose the substrate necessary for vegetation sustenance, triggering reduced fractional vegetation cover and potential ecosystem collapse.

Thus, interaction analysis for this category reveals that under low soil erosion conditions, significant influences are exerted on the SHAP values of six other factors: rocky desertification, elevation, slope gradient, river buffer, runoff buffer, and vegetation coverage. (Figures 13a–f). Specifically, for rocky desertification, elevation, slope gradient, and vegetation coverage, lower feature values correspond to lower SHAP values, while higher feature values yield higher SHAP values under this specific conditional state. (Figures 13a–c,f).

Figure 13

Figure 13. SHAP feature dependency plots for ecological sensitivity influencing factors in slope sensitive category (Class II): (a) Interaction dependency of SHAP values between rocky desertification and soil erosion; (b) Interaction dependency of SHAP values between elevation and soil erosion; (c) Interaction dependency of SHAP values between slope and soil erosion; (d) Interaction dependency of SHAP values between river buffer zone and soil erosion; (e) Interaction dependency of SHAP values between runoff buffer zone and soil erosion; (f) Interaction dependency of SHAP values between vegetation cover and soil erosion.

For Class III - Vegetation Type Sensitive Category, the slope factor shows no significant interactive effect on the SHAP values of the soil erosion factor (Figures 14a). However, the soil erosion factor demonstrates clear interactive influences on both elevation and slope. Specifically, when soil erosion assumes high values, the SHAP values of both elevation and slope exhibit noticeable decreases and increases respectively (Figures 14b,c). When rocky desertification is at low values, it enhances the contribution of the vegetation coverage factor. Under low rocky desertification conditions, higher feature values of vegetation coverage correspond to higher SHAP values, with the maximum value exceeding 0.4. (Figures 14d).

Figure 14

Figure 14. SHAP feature dependency plots for ecological sensitivity influencing factors in vegetation type sensitive category (Class III): (a) Interaction dependency of SHAP values between soil erosion and slope; (b) Interaction dependency of SHAP values between elevation and soil erosion; (c) Interaction dependency of SHAP values between slope and soil erosion; (d) Interaction dependency of SHAP values between vegetation cover and rocky desertification.

Class III areas are characterized as an ecological “critical zone” or “transitional belt.” These areas have not yet fully degraded into barren land (indicated by low rocky desertification values) but remain highly vulnerable. The ecosystem state in these regions critically depends on vegetation resilience and coverage, making vegetation type and coverage the key determinants of sensitivity. The status of vegetation, in turn, interacts with the processes of soil erosion and rocky desertification. Consequently, the interaction between low rocky desertification and high vegetation coverage creates a virtuous cycle. In areas with low rocky desertification threats, every incremental in-crease in vegetation effectively maintains and improves ecosystem health, preventing further degradation toward rocky desertification. Thus, the positive contribution of vegetation coverage becomes substantially magnified.

All the aforementioned interaction characteristics indicate that the ecological sensitivity of this category relies predominantly on vegetation as the crucial element. The greatest diversity of vegetation types in this category suggests high environmental heterogeneity, likely resulting from variations in elevation, aspect, and microhabitats, positioning it within an ecotone where multiple ecological processes intersect. This category simultaneously exhibits both degradation drivers (soil erosion) and restoration potential (the efficient role of vegetation under low rocky desertification conditions). The future trajectory of these ecosystems—whether continuing to degrade or gradually recover—depends on the balance of vegetation-related influencing factors. If vegetation is damaged, soil erosion and slope effects will become dominant, pushing the system toward rocky desertification; if vegetation is protected and restored, its strong positive effects can effectively stabilize the system.

For Class IV - Cultivated Land in River Valleys Sensitive Category, its spatial distribution is highly coincident with cultivated land in the river valley areas of Yangshuo County. This characteristic underscores the profound reshaping of the natural landscape by human agricultural activities. Focusing further on the interaction mechanisms between the land use type factor and other factors reveals that when land use type assumes higher values, it exerts a significant enhancing effect on the SHAP values of three factors: rocky desertification, soil erosion, and aspect. This indicates that agricultural development amplifies the role of these natural factors in determining ecological sensitivity. (Figures 15a–c).

Figure 15

Figure 15. SHAP feature dependency plots for ecological sensitivity influencing factors in cultivated land in river valleys sensitive category (Class IV): (a) Interaction dependency of SHAP values between rocky desertification and land use; (b) Interaction dependency of SHAP values between soil erosion and land use; (c) Interaction dependency of SHAP values between aspect and land use; (d) Interaction dependency of SHAP values between vegetation cover and rocky desertification.

More notably, when the degree of rocky desertification is low, the vegetation coverage factor exhibits a pattern of influence starkly opposite to that observed in the previously analyzed Vegetation Type Sensitive Category (Class III): higher vegetation coverage corresponds to lower SHAP values, meaning it contributes less to classifying a sample into Class IV. Conversely, lower vegetation coverage leads to a significant increase in the SHAP value (exceeding 0.6), making it an important factor driving the identification of samples as this category. (Figures 15d). This unique interactive characteristic aligns well with the practical reality of the river valley cultivated land ecosystem: in these areas, lower vegetation coverage is often associated with intensive crop cultivation or fallow bare states, representing a direct manifestation of agricultural land use. In contrast, higher natural vegetation coverage signifies that the area is undisturbed by agricultural development or has been converted from farmland back to forest, thereby significantly reducing its likelihood of being classified as the Cultivated Land in River Valleys Sensitive Category.

For Class V - Population Aggregation Sensitive Category, the interactions between factors profoundly reveal the adaptive relationship between human settlement selection and the natural environment. When the slope gradient assumes low values (Figures 16a–c), it significantly enhances the contribution of aspect, elevation, and the river buffer zone. Low slope values typically indicate flat, open topography, which represents ideal conditions for constructing towns and transportation infrastructure. Under the premise of population aggregation, the ecological significance of other topographic factors transforms: the importance of aspect increases as it directly affects building daylighting, ventilation, and thermal comfort; the importance of elevation becomes prominent since lower valley elevations not only facilitate easier water access but also correlate with environmental issues like flood risk and temperature inversion (e.g., cold air pooling); the importance of the river buffer zone is amplified because closer proximity to rivers signifies richer water resources but also higher flood disaster risk. Therefore, in flat areas, subtle differences in these factors are magnified by the sensitivity of the human living environment, significantly influencing the probability of an area being classified as Class V.

Figure 16

Figure 16. SHAP feature dependency plots for ecological sensitivity influencing factors in population aggregation sensitive category (Class V): (a) Interaction dependency of SHAP values between aspect and slope; (b) Interaction dependency of SHAP values between elevation and slope; (c) Interaction dependency of SHAP values between river buffer zone and slope; (d) Interaction dependency of SHAP values between soil erosion and land use; (e) Interaction dependency of SHAP values between slope and rocky desertification; (f) Interaction dependency of SHAP values between runoff buffer zone and rocky desertification.

When low land use type values co-occur with low soil erosion characteristic values(Figures 16d), they produce significantly negative SHAP values (<−0.5). This indicates that if an area is both undeveloped (low land use intensity) and exhibits no significant soil erosion, the system determines it does not belong to a population aggregation zone. This is because human settlement activities typically lead to increased land use intensity (e.g., conversion to construction land), disturbance of natural soil, and consequently, some degree of soil erosion. The characteristics of being “undeveloped and ecologically stable” contradict the core features of this “highly anthropogenically disturbed” category, thus strongly inhibiting classification into Class V.

When rocky desertification assumes low values, it markedly enhances the contribution of both slope gradient and the river buffer zone (Figures 16e,f), further corroborating the livability attributes of Class V. Low rocky desertification indicates good surface soil coverage and stable foundations suitable for engineering construction. Against this back-drop of development suitability, the value of two core livability elements—topographic flatness and proximity to rivers—is maximized, collectively rein-forcing the area’s sensitivity and potential as a population aggregation zone.

4 Discussion

4.1 Discussion of research findings

This study systematically identified and analysed the spatial heterogeneity characteristics and dominant influencing factors of ecological sensitivity in Yangshuo County through the “SOM-RF-SHAP” framework. The results indicate that the ecological sensitivity in Yangshuo County can be categorized into five distinct classes, each exhibiting clear geographical distribution patterns and ecological significance. The findings empirically demonstrate that ecological sensitivity in Yangshuo is a composite outcome primarily governed by the “geological-geomorphological” baseline, with localized superimposition of human activities. Spatially, the Vegetation Type Sensitive Category (Class III) and the Cultivated Land in River Valleys Sensitive Category (Class IV) form the dominant pattern, clearly outlining the geographical differentiation where the naturally fragile baseline interweaves with agricultural development activities.

Regarding factor importance, natural baseline factors—lithology, rocky desertification, slope gradient, river buffer zone, soil erosion, and vegetation type—contributed significantly more to ecological sensitivity than anthropogenic and spatial heterogeneity factors, such as land use type, aspect, and elevation. In terms of driving mechanisms, the Random Forest model quantitatively revealed the decisive role of natural baseline factors (lithology, rocky desertification, slope gradient), corroborating the general principle that ecological security in karst areas is constrained by the geological background and pedogenesis processes.

More profound insights emerged from the SHAP interaction analysis, which un-covered differentiated regulatory pathways for various sensitivity categories: In Class III, the positive synergistic effect between low rocky desertification and high vegetation coverage suggests that improving vegetation quality is key to disrupting the positive feedback loop of rocky desertification. In Class IV, the combination of high land use intensity (proportion of cultivated land) and low vegetation coverage strongly indicates a sensitivity type dominated by agricultural disturbance, demonstrating that human activities have surpassed natural factors as the primary controlling factor for ecological sensitivity in these areas. In Class V, the combination of low slope gradient, low rocky desertification, and proximity to rivers significantly enhances habitat suitability, but also simultaneously exposes these areas to dual risks of flooding and development pressure.

The machine learning-driven ecological sensitivity assessment framework developed in this study demonstrates strong potential for extension. On the one hand, although the selected factor set was tailored for the karst environment, the model structure possesses strong indicator compatibility. Future studies could incorporate new indicators such as biodiversity, carbon sink function, and human activity intensity to address policy needs like the “Dual Carbon” goals (carbon peaking and carbon neutrality) or ecological security pattern optimization. On the other hand, this framework has already shown consistent explanatory power in analogous studies of non-karst regions, such as the Chang-Zhu-Tan urban agglomeration, indicating its cross-regional and multi-scale adaptability. The established technical pathway of “unsupervised clustering - machine learning prediction - game theory interpretation” exhibits good universality and extrapolation potential, providing a transferable and reusable analytical tool for ecological sensitivity assessment and spatial regulation in different types of regions.

4.2 Recommendations for future planning and design

The driving mechanisms of ecological sensitivity revealed in this study, particularly the factor interactions quantified by SHAP analysis, provide a scientific basis for differentiated and targeted governance in karst areas. Planning strategies should not remain superficial but must intervene based on the core drivers of different sensitivity categories. Accordingly, the following recommendations for spatial planning and ecological management are proposed. For Class I (Alpine Forest Land Sensitive Category), strict protection should be implemented, limiting development activities. The primary focus should be on conserving primary forest communities and the litter layer to sustainably maintain their water conservation and soil retention functions. Class II (Slope Sensitive Category) is a key zone for soil erosion and rocky desertification occurrence and transmission. The SHAP analysis revealed a cascading positive feedback mechanism of “lithology → slope gradient → soil erosion”. Therefore, management must target this chain: implement the conversion of farmland to forest and grassland on steep slopes outside essential farmland protection boundaries; systematically deploy engineering measures like sediment traps and check dams on slopes, combined with biological measures such as planting drought-tolerant, deep-rooted native shrubs, to directly address the core driver–slope gradient (Wang and Zhang, 2022). For Class III (Vegetation Type Sensitive Category), identified as a priority area for ecological restoration, SHAP interaction analysis indicates a strong positive synergy between “low rocky desertification” and “high vegetation coverage”, signifying that vegetation restoration yields exceptionally high ecological returns. This area should be a priority for ecological restoration projects, but strategies must go beyond simply increasing coverage. Leveraging the high contribution of the vegetation type factor from the SHAP analysis, precise selection of deep-rooted, high-biomass native tree and shrub species (e.g., Cyclobalanopsis glauca, Cinnamomum camphora, Koelreuteria paniculata) for mixed afforestation is crucial (Fang, 2023). The goal should be to establish multi-layered, uneven-aged, multi-functional near-natural forests, thereby efficiently enhancing soil stabilization capacity and blocking the pathway to rocky desertification. For Class IV (Cultivated Land in River Valleys Sensitive Category), the core strategy is balancing agricultural production with ecological protection. This involves vigorously promoting eco-agriculture, converting sloping farmland to economic orchards or implementing contour farming; establishing herbaceous and shrub-based ecological buffer strips in riparian zones to mitigate agricultural non-point source pollution and soil erosion; and exploring fallowing or crop rotation on land with higher ecological value to improve overall regional vegetation coverage. Slope farmland management and the construction of riparian ecological interception systems should be strengthened (Luo et al., 2022). For Class V (Population Aggregation Sensitive Category), ecological adaptation principles should be embedded in urban construction. This includes controlling disordered development within river buffer zones, considering flood control and drainage capacity under extreme rainfall, and constructing sponge city facilities while preserving permeable areas to reduce waterlogging risk.

Given the core roles of lithology, slope gradient, and rocky desertification, it is recommended to implement zoning regulations based on geological and topographic constraints within territorial spatial planning, strictly limiting construction activities on steep slopes and areas with highly sensitive lithology. Vegetation type and coverage are key adjustable factors; ecological engineering should focus on improving vegetation quality, not just quantity, promoting near-natural vegetation restoration, especially in ecotones. Although human activity factors like land use show lower overall importance, their impact is significant in localized areas such as valley farmland and urban peripheries. It is advised to strictly adhere to ecological protection redlines during development, coordinating the relationship between agriculture, urban development, and ecological conservation.

4.3 Study limitations

This study, while attempting to enhance the objectivity and interpretative depth of ecological sensitivity assessment using machine learning methods, has several limitations. Some of the fundamental data used, such as soil type, lithology distribution, and vegetation coverage, were primarily sourced from public datasets or medium-resolution remote sensing imagery. Constrained by the precision and update frequency of the original data, potential biases exist in representing fine-scale geomorphic units or areas subject to short-term anthropogenic disturbances. On the data level, some of the fundamental data used in this study are constrained by the precision of publicly available datasets, which may not fully capture the characteristics of fine-scale geomorphic units. More importantly, several key factors unique to karst ecosystems, such as “karst fissure density” and “spatial distribution of underground rivers,” were not incorporated into the modeling due to the challenges in obtaining such data over large areas. These factors directly control the exchange between surface water and groundwater, soil retention capacity, and nutrient transport. Their absence may limit the model’s ability to fully characterize the ecological effects of the surface-subsurface dual hydrological structure in karst systems and could lead to biases in the assessment of ecological sensitivity in certain regions, such as areas strongly influenced by underground rivers. Future research could leverage higher-precision geological surveys and geophysical exploration techniques to acquire these critical parameters.

Although the Random Forest and SHAP methods significantly enhanced the interpretation of factor contributions and interactions, they remain essentially explanatory strategies and cannot fully reconstruct definitive causal mechanisms between factors. The model results primarily reflect statistical associations rather than being driven by explicitly modeled ecological processes. For instance, while the SHAP values indicated a strong interaction between rocky desertification and vegetation coverage, the underlying intermediate processes—such as soil thickness and water stress—require further validation through field observations and mechanistic modeling. In Regarding scale, the findings of this study are based on the county level of Yangshuo County. Validation across multiple scales (e.g., township, small watershed) has not been conducted, nor has a systematic comparison been made with broader karst regions. The applicability, stability, and parameter transferability of the model and factor system across different spatial granularities or heterogeneous geographical contexts require further testing. The importance ranking and interaction characteristics of certain factors might differ in analyses conducted at municipal, county, or village levels.

This assessment, based on data from a current period, represents a static identification of sensitivity. It did not incorporate time-series remote sensing and socio-economic data to reveal the dynamic evolution patterns and driving mechanisms of sensitivity. Given the significant temporal accumulation and hysteresis effects inherent in ecological sensitivity and rocky desertification processes in karst areas, future research could introduce time-series modeling to enhance the predictive capability regarding sensitivity evolution trends. Furthermore, although the current factor selection covers the natural baseline and some anthropogenic pressures, management-oriented indicators—such as those related to ecological restoration projects, protected area planning, and socio-economic activities—were not sufficiently incorporated. This limitation somewhat restricts the direct applicability of the evaluation results to decision-making needs like territorial spatial planning and site selection for ecological restoration projects. Subsequent studies could attempt to introduce more quantifiable and spatializable governance-related variables to strengthen the policy support capacity of the findings.

5 Conclusion

This study focused on Yangshuo County, Guilin—a typical karst area—and established an integrated ecological sensitivity assessment framework by introducing the Self-Organizing Map (SOM) for clustering, the Random Forest (RF) machine learning algorithm, and the SHAP interpretability model. This framework combines objective classification, factor importance evaluation, and the analysis of multi-factor interaction mechanisms. The results demonstrate significant spatial heterogeneity in ecological sensitivity within Yangshuo County, which can be categorized into five distinct types. Among these, Category III (Vegetation Type Sensitive) and Category IV (Cultivated Land in River Valleys Sensitive) dominate the landscape, reflecting the combined effects of an inherently fragile natural baseline and human activity disturbances. The Random Forest model identified natural baseline factors—lithology, rocky desertification, and slope gradient—as the dominant drivers of sensitivity, whose importance significantly surpasses that of anthropogenic factors. This finding corroborates the fundamental principle that ecological sensitivity in karst regions is governed by its “geology-geomorphology” foundation. SHAP interaction analysis further revealed non-linear synergistic mechanisms among multiple factors. For instance, the combination of low rocky desertification and high vegetation coverage significantly enhances ecological stability, whereas high land use intensity markedly increases ecological risk in valley cultivated lands. The methodology employed in this study effectively overcomes the limitations of traditional methods like AHP and PCA, particularly regarding subjective weighting and linear assumptions. It provides a transferable and interpretable analytical paradigm for ecological sensitivity assessment in karst regions. The findings offer a scientific basis for ecological conservation, restoration, and territorial spatial planning in Yangshuo County and similar areas. Future research should focus on multi-scale validation, dynamic simulation, and the integration of policy-related factors to further enhance the model’s predictive capability and decision-support potential.

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Author contributions

Y-tZ: Investigation, Writing – original draft, Resources, Visualization, Formal Analysis, Validation, Data curation, Methodology, Conceptualization, Writing – review and editing. Y-zW: Writing – original draft, Formal Analysis, Visualization. X-lL: Formal Analysis, Writing – original draft, Software, Data curation. YS: Writing – original draft, Validation, Methodology. JY: Project administration, Writing – review and editing, Supervision, Funding acquisition.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This research was funded by “The Fundamental Research Funds for the Provincial Universities of Liaoning, grant number LJ142510152001”.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Breiman, L. (2001). Random forests. Mach. Learn. 45 (1), 5–32. doi:10.1023/a:1010933404324

CrossRef Full Text | Google Scholar

Chang, J., Li, Q., Zhai, L., Liao, C., Qi, X., Zhang, Y., et al. (2024). Comprehensive assessment of rocky desertification treatment in southwest China karst. Land Degrad. and Dev. 35 (10), 3461–3476. doi:10.1002/ldr.5146

CrossRef Full Text | Google Scholar

Chen, Q. (2024). Research on the main problems and countermeasures of ecological environment governance in the li river yangshuo basin under the goal of building a world-class tourist city. Guilin: Guangxi Normal University Press.

Google Scholar

Chen, Y., Cheng, C., Xiong, K., Rong, L., and Zhang, S. (2024). Quantifying the biodiversity and ecosystem service outcomes of karst ecological restoration: a meta-analysis of south China karst. CATENA 245, 108278. doi:10.1016/j.catena.2024.108278

CrossRef Full Text | Google Scholar

Chen, Y., Xie, B., and Li, J. (2025). Ecological sensitivity and driving factors in the chang-zhu-tan urban agglomeration based on machine learning. Ecol. Sci. 44 (01), 28–39. doi:10.14108/j.cnki.1008-8873.2025.01.004

CrossRef Full Text | Google Scholar

Chen, L., Chen, R., and Shi, Y. (2025). GIS-based study on ecological environment sensitivity in Ningxia. Geomatics Technol. Equip. 27 (03), 1–6. doi:10.20006/j.cnki.61-1363/P.2025.03.001

CrossRef Full Text | Google Scholar

D'Ettorre, U. S., Liso, I. S., and Parise, M. (2024). Desertification in karst areas: a review. Earth-Science Rev. 253, 104786. doi:10.1016/j.earscirev.2024.104786

CrossRef Full Text | Google Scholar

Dao-Xian, Y. (2008). Global view on karst rock desertification and integrating control measures and experiences of China. Pratacultural Sci. doi:10.1007/s10499-007-9164-4

CrossRef Full Text | Google Scholar

Dong, W., Su, W., and Gou, R. (2025). Comprehensive assessment and spatiotemporal evolution of ecological sensitivity in a karst Mountain city. Res. Soil Water Conservation 32 (02), 276–285. doi:10.13869/j.cnki.rswc.2025.02.001

CrossRef Full Text | Google Scholar

Eker, R., and AydıN, A. (2024). Predicting potential fire severity in Türkiye’s diverse forested areas: a SHAP-Integrated random forest classification approach. Stoch. Environ. Res. Risk Assess. 38 (12), 4607–4628. doi:10.1007/s00477-024-02820-1

CrossRef Full Text | Google Scholar

Fan, Q. (2023). Construction and application of a karst disturbance index evaluation system. Guilin: Guangxi Normal University Press.

Google Scholar

Fang, J. (2023). Regulation of grassland ecosystem vulnerability and strategies for enhancing ecological resilience in karst rocky desertification control. Guiyang: Guizhou Normal University Press.

Google Scholar

Gong, Y. (2023). Zoning of territorial space ecological restoration in karst areas guided by landscape value conservation. Nanning: Guangxi University Press.

Google Scholar

Ji, D., Li, S., and Sun, J. (2023). Land ecological sensitivity assessment in yanbian area of Jilin Province based on random forest algorithm. Glob. Geol. 42 (04), 761–767.

Google Scholar

Jia, H., Ma, X., and Wang, W. (2024). Spatial identification for ecological protection and restoration in a typical karst area: a case study of shilin yi autonomous county. J. Yunnan Agric. Univ. Nat. Sci. 39 (02), 145–154. doi:10.12101/j.issn.1004-390X(n).202305006

CrossRef Full Text | Google Scholar

Jiang, L. (2024). Research on efficient flood risk prediction based on machine learning. Guangzhou: South China University of Technology Press.

Google Scholar

Kohonen, T. (2013). Essentials of the self-organizing map. Neural Netw. 37, 52–65. doi:10.1016/j.neunet.2012.09.018

PubMed Abstract | CrossRef Full Text | Google Scholar

Li, Z., Yin, X., and Liu, Y. (2024). Spatiotemporal evolution of rocky desertification and ecological sensitivity assessment in a typical karst Mountain area of southeastern Yunnan. Ecol. Environ. Sci. 33 (11), 1792–1802. doi:10.16258/j.cnki.1674-5906.2024.11.013

CrossRef Full Text | Google Scholar

Liang, Y.-X., Wu, D.-F., Wu, Z.-J., Xu, Y., Zhu, Z. W., Zhang, Y. C., et al. (2023). Construction of ecological corridors in karst areas based on ecological sensitivity and ecological service value. Land (Basel). 12 (6), 1177. doi:10.3390/land12061177

CrossRef Full Text | Google Scholar

Lin, L. (2024). Ecological sensitivity assessment in national key ecological function areas: a case study of jianhe county, Guizhou province. Agric. Technol. 44 (16), 97–102. doi:10.19754/j.nyyjs.20240830021

CrossRef Full Text | Google Scholar

Liu, Y. (2024). Construction of ecological security pattern in the karst Mountain area of southeastern Yunnan. Kunming: Yunnan University of Finance and Economics Press.

Google Scholar

Liu, J., and Li, Y. (2012). Ecological security assessment of rocky desertification area in Western Hunan Province based on PCA method. J. Central South Univ. Sci. Technol. 43 (12), 4895–4901.

Google Scholar

Lu, X. (2024). A review of karst environment and biodiversity in China. Advances in Environmental Protection.

Google Scholar

Lundberg, S. M., and Lee, S.-I. (2017). “A unified approach to interpreting model predictions,” in Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, California, USA (Long Beach, CA: Curran Associates Inc.), 4768–4777.

Google Scholar

Luo, Y., Newman, C., Luo, Y., and Zhou, Z. M. (2022). Investigation and evaluation of reserve cultivated land resources based on GIS technology: a case study of yangshuo county. South. Nat. Resour. 13 (06), 40–45+53. doi:10.3390/ani13010040

PubMed Abstract | CrossRef Full Text | Google Scholar

Qiu, S., Hu, J., and Yang, Z. (2024). Construction of ecological security pattern in li river basin based on ecosystem services and ecological sensitivity assessment. J. Northwest For. Univ. 39 (06), 153–162.

Google Scholar

Tian, Y., and Wang, Y. (2025). GIS and AHP-based ecological sensitivity assessment in the li river basin. South. Nat. Resour. (08), 23–28.

Google Scholar

Tu, C., Luo, W., and Chen, Y. (2024). Zonal governance of karst landscape resources in guilin based on ecosystem sensitivity and service function. Chin. Geol. 51 (06), 1839–1854. doi:10.12029/gc20230906001

CrossRef Full Text | Google Scholar

Wang, X., and Shang, Z. (2025). Analysis of county-scale habitat restoration strategies based on ecological sensitivity: a case study of wanrong county. J. Green Sci. Technol. 27 (11), 215–222. doi:10.16663/j.cnki.lskj.2025.11.002

CrossRef Full Text | Google Scholar

Wang, B., and Zhang, L. (2022). Meteorological early warning for slope geological hazards in yangshuo county. West-China Explor. Eng. 34 (06), 19–22.

Google Scholar

Wang, Y., and Long, L. (2024). “Research on supply and demand of ecosystem services and ecological restoration in a karst city: a case study of guilin [C],” in Proceedings of the 2024 China Urban Planning Annual Conference, Hefei, Anhui, China.

Google Scholar

Wei, D. (2024). GIS-Based assessment of geological disaster susceptibility in yangshuo. Guilin: Guilin University of Technology Press.

Google Scholar

Yang, H., Song, X., and Wang, Y. (2025). GIS-based ecological sensitivity assessment at the county scale: case studies of aksai county and subei county in Gansu province. Geol. Resour. 34 (04), 479–488. doi:10.13686/j.cnki.dzyzy.2025.04.010

CrossRef Full Text | Google Scholar

Ying, B., Liu, T., Ke, L., Xiong, K., Li, S., Sun, R., et al. (2023). Identifying the landscape security pattern in karst rocky desertification area based on ecosystem services and ecological sensitivity: a case study of guanling county, Guizhou province. Forests 14 (3), 613. doi:10.3390/f14030613

CrossRef Full Text | Google Scholar

Yuan, Q. (2025). Ecological sensitivity assessment and landscape restoration strategies in the karst region of southern China. Guiyang: Guizhou Normal University Press.

Google Scholar

Zhang, J. (2023). Study on the synergistic mechanism and model between conservation of karst world heritage site and tourism development in its buffer zone. Guiyang: Guizhou Normal University Press.

Google Scholar

Zhang, Z., Wang, C., and Lv, B. (2024). Comparative analysis of ecological sensitivity assessment using the coefficient of variation method and machine learning. Environ. Monit. Assess. 196 (10), 1000. doi:10.1007/s10661-024-13195-9

PubMed Abstract | CrossRef Full Text | Google Scholar

Zhu, B. (2021). Assessment of ecosystem service functions in a typical karst peak-cluster depression in guilin using the InVEST model. Beijing: Chinese Academy of Geological Sciences Press.

Google Scholar

Zhu, X., Wang, J., and Fu, J. (2023). “Health assessment of small and medium-sized Rivers in karst areas,” in Proceedings of the 2023 China Hydraulic Engineering Academic Conference, Zhengzhou, Henan, China.

Google Scholar