Effects of mapping units on landslide susceptibility assessment: a comparative analysis of slope units and raster cells

1 Introduction

Slope failures rank among the most widespread and destructive geological hazards, continually endangering human life, assets, engineering facilities, and regional socioeconomic stability (Shen et al., 2025; Baral et al., 2021). In China, complex topography, active tectonics, and monsoon rainfall have contributed to frequent landslides in recent years. This risk has been further intensified by more frequent extreme rainfall events and increasing human engineering activities, resulting in widespread landslide hazards. In this context, accurately identifying regional high-susceptibility areas is essential for improving geological-disaster prevention and supporting territorial spatial planning (Halder et al., 2025).

To systematically reveal the spatial distribution patterns and potential risks of landslide disasters at the regional scale, numerous studies have adopted Landslide Susceptibility Assessment (LSA) methods (Ma et al., 2025; Chen et al., 2025a). LSA provides a quantitative estimation of the relative susceptibility of mapping units by integrating multi-source conditioning factors, such as terrain characteristics, geological settings, hydrometeorological conditions, and anthropogenic influences. Accordingly, LSA serves as an essential foundation for landslide hazard evaluation, risk-based spatial zoning, and mitigation-oriented disaster management, making it an integral part of contemporary landslide studies (Chen et al., 2025b; Abay et al., 2025; Goetz et al., 2015).

In LSA modeling, mapping-unit selection is a key step that links environmental conditioning factors to landslide occurrence records. It determines the spatial representation of environmental information and sample organization, and directly affects computational efficiency, physical interpretability, and predictive accuracy (Huang et al., 2024; Rolain et al., 2023). A variety of mapping units have been adopted in LSA, such as raster cells, slope units, watershed-based units, and geomorphological units, with raster cells and slope units being the two most commonly used schemes. Raster-based approaches are computationally efficient, easy to implement, and compatible with multi-source raster datasets; therefore, they have been widely used in regional-scale susceptibility studies. In contrast, slope units have boundaries that better conform to natural slope morphology and provide clearer geomorphological and engineering-geological meaning. They can better represent the spatial homogeneity of controlling factors such as topography and lithology, thereby providing a stronger physical basis for susceptibility assessment (Ma et al., 2023; Chang et al., 2023).

However, both mapping-unit types have inherent limitations. Raster cells often fragment the integrity of natural slopes and fail to fully capture geomorphological units governing landslide initiation and evolution. In complex mountainous regions, scale and boundary effects may introduce additional uncertainties. Conversely, slope units rely on digital elevation models (DEM) and hydrological segmentation algorithms, making their delineation sensitive to DEM resolution, preprocessing procedures, and parameter thresholds. Different delineation schemes may lead to significant variations, and the internal heterogeneity of slope units may weaken their assumption of spatial uniformity (Alvioli et al., 2016). At present, no consensus has been reached on the most suitable mapping unit for LSA. For instance, Li et al. (2025) reported that slope units outperform raster cells in terms of factor response, model discrimination ability, and early warning accuracy in Qingchuan County, Sichuan Province. In contrast, Xu et al. (2024) found that a random forest model based on raster cells achieved higher predictive accuracy in the Zagunao River basin of Li County. These discrepancies suggest that the optimal mapping unit is highly region-dependent and may be influenced by local geomorphology, geological structure, landslide type, and data resolution. Moreover, Li et al. (2022) further demonstrated that even within slope-unit frameworks, different delineation strategies (e.g., hydrological division versus curvature-based watershed segmentation) can alter spatial representation and thus affect susceptibility outcomes.

Although previous studies have compared raster cells and slope units in LSA, existing comparisons have primarily focused on differences in predictive accuracy, susceptibility classification, or the effects of unit delineation. Consequently, it remains insufficiently understood whether mapping-unit effects are consistent across models with different learning structures and, more importantly, how mapping-unit representation affects inter-model consistency, spatial susceptibility patterns, and the way predictive models use environmental information. Therefore, mapping-unit effects need to be examined beyond predictive accuracy by considering both model outputs and the underlying decision mechanisms.

Driven by ongoing methodological development, LSA studies have gradually moved beyond expert-based evaluation and conventional statistical techniques toward data-driven machine-learning approaches (Abdelkader and Csámer, 2025). In recent years, machine-learning models have been widely adopted in LSA because of their ability to characterize complex nonlinear associations and process high-dimensional data (Zhang et al., 2019; Gao et al., 2022). However, although these black-box models can improve predictive performance, their limited interpretability makes it difficult to quantify the contributions of key conditioning factors. This limitation restricts their practical application in disaster-mitigation decision-making. Studies such as Ma et al. (2024) have shown that introducing explainable AI methods such as SHAP can effectively reveal model decision logic and factor contribution mechanisms. In addition, LSA results are inherently subject to multiple sources of uncertainty, including model selection, parameter tuning, sampling strategy, and data quality. Although multiple performance metrics are now commonly used to assess predictive performance, the effects of mapping-unit selection on spatial uncertainty and inter-model consistency still warrant further investigation (Chen et al., 2026; Huang et al., 2024).

The G0711 Urumqi–Wuwei Expressway is a key transportation infrastructure project in the Xinjiang Uygur Autonomous Region of China. Located in a complex mountain–basin transitional environment, the route is exposed to significant geological hazard risks during both construction and operation. Hejing County, a critical area along this transportation corridor, is characterized by diverse landforms, strong topographic relief, and complex slope structures and external disturbance conditions, resulting in prominent landslide susceptibility. The G0711 Urumqi–Wuwei Expressway is a key transportation infrastructure project in the Xinjiang Uygur Autonomous Region of China. Hejing County, located in a complex mountain–basin transitional environment along this transportation corridor, was selected as the study area because of its pronounced topographic relief, diverse geomorphological conditions, and significant engineering disturbance. Based on a unified conditioning-factor database and landslide inventory, this study constructed two mapping-unit schemes, namely, 30 m raster cells and slope units, and applied four representative machine-learning algorithms (LR, MLP, RF, and XGBoost) under both schemes.

Rather than simply determining which mapping unit achieves higher predictive accuracy, this study examines mapping-unit effects from three complementary perspectives: (1) whether the effects of mapping-unit selection are consistent across machine-learning models with different structures and how they influence inter-model variability; (2) how raster cells and slope units reshape the statistical distribution and spatial representation of landslide susceptibility; and (3) whether mapping-unit representation alters model dependence on environmental conditioning factors, as revealed through SHAP analysis. By integrating predictive performance, inter-model consistency, spatial representation, and model interpretation within a unified comparative framework, this study aims to provide a more process-oriented understanding of mapping-unit effects and a practical basis for selecting mapping units according to the spatial scale and objective of LSA.

2 Methodology

2.1 Technical workflow

To examine how mapping-unit schemes affect landslide susceptibility mapping outputs and predictive modeling performance, the detailed workflow illustrated in Figure 1 was summarized into four main stages: data acquisition and conditioning-factor preparation, mapping-unit construction, feature extraction and model training, and result evaluation and comparative analysis.

• Based on multi-source datasets, including a historical landslide inventory, DEM, geological maps, precipitation data, road networks, river systems, and NDVI, conditioning factors were selected according to landslide-formation mechanisms: topography and geomorphology, land cover, hydrological conditions, geological conditions, rainfall conditions, and anthropogenic activities. Spatial alignment, factor calculation, and normalization were then performed to ensure data consistency.
• On this basis, two mapping-unit systems were constructed: regular raster cells and slope units. Raster cells were constructed through systematic grid-based subdivision, while slope units were extracted using a hydrology-driven delineation procedure. For each mapping-unit framework, landslide susceptibility prediction (LSP) was implemented with four machine-learning models, namely, LR, MLP, RF, and XGBoost.
• Lastly, the generated model results were assessed and benchmarked based on their prediction performance, SHAP-based feature interpretation, susceptibility-index distribution, and spatial-pattern differences. This framework enabled a systematic assessment of the applicability of the two mapping-unit types and their effects on model performance and spatial representation.

2.2 Mapping unit construction

Mapping-unit selection is a critical step in LSA because it links spatially continuous environmental conditioning factors with discrete landslide samples and determines the spatial representation and statistical properties of model inputs (Liu et al., 2025; Yang et al., 2025). In this study, the spatial resolution of the raster-based units was uniformly set to 30 m. Slope units were delineated using the r.slopeunits module implemented in GRASS GIS. The slope-unit delineation procedure is illustrated in Figure 2. Because slope-unit geometry is sensitive to parameter settings, several combinations of the controlling parameters were tested according to the terrain relief, drainage-network characteristics, and spatial scale of Hejing County. The final parameter combination was selected by visually and geomorphologically examining the size, shape, and correspondence of the resulting units with terrain boundaries.

The initial flow-accumulation area threshold (thresh) was finally set to 1,000,000 m², the minimum slope-unit area (areamin) to 50,000 m², and the minimum circular-variance threshold (cvmin) to 0.3. The threshold reduction factor was set to 10, and the maximum number of iterations (maxiteration) was limited to 3. The main procedures were as follows:

• Flow-direction and flow-accumulation rasters were derived from the DEM to identify potential terrain boundaries and drainage divides.
• Threshold-based segmentation was applied to generate an initial channel network and preliminary terrain partitions.
• The initial units were subsequently merged and refined based on the circular variance of slope aspect and local topographic homogeneity to ensure internal consistency within slope units.
• An iterative optimization procedure was performed to merge excessively small or morphologically unreasonable units according to the predefined area and circular-variance criteria.
• After automated delineation and iterative optimization, the resulting slope units were imported into ArcGIS for manual inspection against the DEM, hillshade, drainage lines, and ridge–valley structures. Where automatically generated boundaries showed obvious deviations from the actual topographic divides or slope morphology, local boundary corrections and unit adjustments were performed. The final slope-unit dataset used for subsequent LSA modeling therefore represents the result of automated delineation, parameter optimization, and manual geomorphological refinement.

2.3 Prediction models

2.3.1 Logistic regression

LR is a generalized linear model commonly used for binary classification. It estimates the effects of environmental conditioning factors on the log-odds of landslide occurrence using maximum likelihood estimation (Puente-Sotomayor et al., 2021). The linear predictor was transformed into a probability of landslide occurrence using the sigmoid function, yielding a susceptibility value ranging from 0 to 1. The mathematical formulation of the model is given in Equations 1, 2: where p denotes the predicted probability of landslide occurrence, constrained to range from 0 to 1; z denotes the logit-scale predictor derived by combining the conditioning variables with their assigned weights; x _*m* corresponds to the m-th environmental factor; b _*m* is its associated model coefficient; and a represents the constant term.

z = a + ∑ 1 m b m x m (1)

p = 1 1 + e − z (2)

2.3.2 Multilayer perceptron

As a conventional feedforward neural network, MLP is widely employed to address both categorical prediction tasks and regression-based modeling problems. Structurally, an MLP consists of multiple successive layers, typically comprising an input stage, one or more intermediate hidden layers, and a final output layer, with neurons in neighboring layers connected in a dense manner (Chang et al., 2023). Nonlinear activation functions are applied to all computational neurons, while the input nodes serve only to pass the original features into the network.

In the present analysis, the normalized conditioning variables served as input features, whereas the final layer produced the corresponding landslide-susceptibility probability. The hidden layers were responsible for extracting nonlinear feature interactions, and the inter-neuron weights were iteratively adjusted during model training to reduce predictive loss. The hidden layers were responsible for extracting nonlinear feature interactions, and the inter-neuron weights were iteratively adjusted during model training to reduce predictive loss. The hidden-layer response is defined in Equation 3:where Y refers to the hidden-layer output, X denotes the input feature vector, W represents the matrix of connection weights, b is the bias parameter, and σ indicates the activation function. For the MLP implementation, sigmoid-based nonlinear transformation was applied throughout the model development and validation phases.

Y = σ W X + b (3)

2.3.3 Random forest

RF performs classification by aggregating predictions from multiple tree-structured learners within an ensemble framework. Its predictive capability is strengthened by combining bootstrap resampling, stochastic feature selection, and aggregated tree outputs (Qiao and Chang, 2021). In LSA modeling, RF constructs multiple bootstrap training subsets from the original dataset, with each subset having the same sample size as the original data. Each subset is used to train an individual base learner, and the ensemble model is formed accordingly.

For each subset, an independent decision tree is constructed. In this study, the RF model used all 12 conditioning factors as input features (m = 12), and three features (n = 3) were randomly considered as candidate variables at each node split according to the optimized max_features setting. After each decision tree outputs the landslide-occurrence probability for a given mapping unit, the RF model aggregates the predictions of all trees by averaging, yielding a continuous susceptibility score bounded between 0 and 1, as shown in Equation 4.where I is the set of all classes, k is the number of decision trees, p _*i* is the probability of event i, p _*ij* is the probability of event i predicted by the jth decision tree, and P is the final probability corresponding to the selected class.

P = max p i = ∑ j = 1 k p i j k | i ∈ I (4)

2.3.4 XGBoost

XGBoost constructs a strong predictive learner by sequentially combining multiple weak models under a gradient-boosting scheme. It improves predictive performance by sequentially constructing decision trees designed to compensate for residual errors from earlier learning rounds (Hong et al., 2022). In LSA modeling, XGBoost minimizes an objective function based on prediction errors and regularization terms. At each boosting round, an additional decision tree is incorporated to update and improve the current predictive function. Tree construction is guided by an objective function that uses both first- and second-order gradient information, while regularization is applied to constrain model complexity and reduce the risk of overfitting. The tree-level outputs were subsequently integrated to derive the final landslide-susceptibility likelihood for each mapping unit, as expressed in Equation 5.where T represents the number of trees in the ensemble; f _*t* (x _*i*) refers to the output assigned to sample x _*i* by the tth decision tree; σ(.) refers to the sigmoid transformation, which converts the aggregated model output into a probability value within the interval [0, 1]; and P denotes the final landslide susceptibility probability for each sample unit.

P = σ ∑ t = 1 T f t x i (5)

2.4 Evaluation and interpretation of LSA results

2.4.1 Accuracy evaluation metrics

Model performance was assessed using a set of evaluation metrics, including accuracy (R _*A*), precision (R _*P*), recall (R _*R*), F1 score (S _*F1*), and the receiver operating characteristic (ROC) curve, to provide a comprehensive appraisal of the predictive reliability of the LSA models (Chicco et al., 2021).

The ROC curve is derived from confusion-matrix statistics by plotting the false positive rate (R _*FP*) along the horizontal axis against the true positive rate (R _*TP*) along the vertical axis. It is used to examine model classification behavior across varying threshold settings. When the ROC curve approaches the upper-left region of the plot, the model is considered to have stronger discrimination capability. The area under the curve (AUC) further provides a numerical summary of overall predictive performance, ranging from 0 to 1, where larger values reflect better model effectiveness.

Equations 6–10 present the mathematical definitions of the performance-evaluation indicators: where PT refers to the true-positive count, representing landslide samples that are correctly identified; NT indicates the true-negative count, corresponding to non-landslide samples that are correctly recognized; PF indicates non-landslide cases that were erroneously labeled as landslides, whereas NF refers to landslide cases that were mistakenly assigned to the non-landslide class.

R A = P T + N T P T + P F + N T + N F (6)

R P = P T P T + P F (7)

R R = R T P = P T P T + N F (8)

S F 1 = 2 × R P × R R R P + R R (9)

R F P = P F P F + N T (10)

2.4.2 Mean and standard deviation of LSI

The central tendency and variability of the landslide susceptibility index (LSI) are characterized by its mean value and standard deviation, respectively. Using these two statistical measures to characterize the distribution of LSI values can clarify how mapping-unit choices affect model-specific LSA outputs. Generally, smaller mean values combined with larger standard deviations suggest that the model has greater capacity to differentiate spatial variations in landslide susceptibility (Huang et al., 2025).

2.4.3 SHAP-based model interpretation

SHAP provides a feature-attribution approach for elucidating the decision behavior of predictive models. It is based on the Shapley value allocation principle, which decomposes model predictions into contributions from individual input features and quantifies the influence of each variable on model decisions (Ji et al., 2023). This method represents model output as an additive combination of feature-contribution values, enabling the interpretation of complex machine-learning models.

Within the LSA framework, SHAP enables the contribution of conditioning factors to landslide-occurrence probability to be interpreted at both the overall model level and the individual-sample level, while also indicating each factor’s effect direction and relative contribution. The underlying formulation is given in Equation 11:where y _*i* represents the model prediction for sample i; f (x _*ij*) denotes the SHAP contribution assigned to the jth feature of that sample. A positive value of f (x _*ij*) (f (x_{ij})) indicates that the feature raises the predicted outcome, whereas a negative value suggests that the feature suppresses the prediction.

y i = y base + f x i 1 + f x i 2 + ⋯ + f x i k (11)

3 Study area and data sources

3.1 Study area overview

Hejing County lies in northern Bayingolin Mongolian Autonomous Prefecture, Xinjiang Uygur Autonomous Region, China. Geographically, it occupies the transitional region from the southern piedmont of the central Tianshan Mountains to the northern edge of the Tarim Basin, spanning approximately 82°28′–87°52′E and 42°06′–43°33′N. The county covers approximately 33,799.56 km² and is one of the largest county-level administrative units in Xinjiang.

The study area is situated in a typical mountain-basin transition zone of the Tianshan Mountains and consists of high mountains, alluvial fans, and alluvial-diluvial plains, representing characteristic arid-region mountainous geomorphology (Figure 3). Geologically, the area belongs to the North Tianshan tectonic belt of the Tianshan fold system and is strongly controlled by the North Tianshan fault system and thrust-nappe structures, reflecting intense tectonic activity.

The stratigraphy in the region is well exposed, including Paleozoic metamorphic and sedimentary rocks, Mesozoic clastic rocks, and Cenozoic unconsolidated sediments. The lithological assemblages are complex and highly heterogeneous. Multiple phases of tectonic activity have resulted in densely developed fault systems, and rock masses in piedmont and valley areas are highly fractured, providing favorable geological conditions for landslide development.

The terrain generally descends from the northern highlands toward the southern lowlands and can be classified into glaciated and snow-covered high-mountain areas, mountainous erosional zones, and piedmont alluvial–diluvial plains. The mountainous areas are characterized by steep slopes and deep valleys with strong relief. Under the influence of heavy rainfall and snowmelt runoff, unstable slopes are prone to landslides and collapses. The landslide inventory compiled for this study includes 128 documented landslide occurrences.

3.2 Factor database

3.2.1 Data sources

Spatial analysis and model implementation were performed using ArcGIS 10.8 and PyCharm, respectively. The dataset includes both basic geographic information and thematic environmental variables. The basic datasets consist of a 30 m resolution DEM and transportation and geological structure maps at a scale of 1:250,000. The thematic datasets integrate vegetation coverage, annual precipitation, river networks, road systems, and landslide inventory points across the investigated region.

To maintain spatial alignment among the input layers, all datasets were transformed to the Xi’an 1980 coordinate reference system. The corresponding data characteristics and source information are provided in Table 1.

TABLE 1. Data sources and data types.

3.2.2 Selection of conditioning factors

Based on the geomorphological characteristics and landslide-triggering mechanisms of Hejing County, this study constructed an LSA indicator system that includes topographic, geomorphological, land-cover, hydrological, geological, rainfall, and anthropogenic factors. Twelve conditioning factors were incorporated in total (Figure 4), as described below.

• Topographic and geomorphological factors

Slope, aspect, elevation, relief amplitude, and plan curvature were extracted from the digital elevation model (DEM). Among these variables, slope controls surface runoff, elevation reflects gravitational potential, aspect influences local hydrothermal conditions, and relief and curvature characterize terrain morphology.

• 2. Land cover factors

NDVI, distance to roads (RoadDis), and terrain roughness (TerrRough) were selected. NDVI represents vegetation coverage, RoadDis reflects the intensity of human disturbance, and TerrRough describes micro-topographic variability.

• 3. Hydrological factors

The hydrological conditioning variables comprise precipitation, the stream power index (SPI), the topographic wetness index (TWI), and river proximity (RiverDis). These variables characterize precipitation input, erosional potential, flow-accumulation capacity, and terrain moisture conditions.

• 4. Lithological factor

Lithology governs the strength characteristics and permeability behavior of rock–soil materials. The study area is mainly composed of Paleozoic metamorphic rocks, sedimentary rocks, and Mesozoic clastic rocks, with locally developed Cenozoic unconsolidated deposits widely distributed in piedmont alluvial–diluvial plains and river valleys. These lithologies are generalized into four categories: hard rock (HR), interbedded hard–soft rock (IHSR), soft rock (SR), and unconsolidated deposits (UD). As lithology is represented as a categorical geological variable, spatially adjacent mapping cells located within the same geological unit are expected to share the same lithological category.

3.2.3 Conditioning factor processing

In this study, an identical set of conditioning factors is used for both raster-based and slope-unit-based modeling. However, because the two mapping-unit types differ in spatial scale and representation, the conditioning factors were processed differently.

• At the raster-cell scale, units are small and regular, and each cell value represents local environmental attributes. To eliminate differences in units and value ranges among variables, all conditioning factors were normalized to a common numerical range, thereby improving model-training stability and comparability.
• At the slope-unit scale, units have clear drainage boundaries and larger spatial extents. Accordingly, raster-derived conditioning factors were first summarized at the slope-unit level, and the average value of each factor was assigned to characterize the corresponding slope unit. The resulting mean values are then normalized to remove dimensional effects, producing standardized model inputs.

Following this procedure, all conditioning factors were converted into dimensionless features, thereby ensuring uniformity and comparability among the model input variables. Given that the present analysis emphasizes comparisons among mapping-unit designs rather than differences in input-variable magnitudes, this unified framework enables a systematic investigation of how mapping unit type and spatial scale influence LSA results.

3.3 Mapping unit delineation

In this study, the raster-cell resolution was set to 30 m, resulting in 75,697,617 raster cells. Using the hydrology-based delineation procedure described previously, a total of 73,131 slope units were generated, as illustrated in Figure 5. The areas of the slope units ranged from 0.035 to 8.985 km², with a mean of 0.474 km² and a standard deviation of 0.440 km². The slope units were partitioned along natural drainage divides and slope morphological boundaries to preserve the structural integrity and topographic continuity of individual slopes as much as possible (Woodard et al., 2024).

4 Results and analysis

4.1 Model training and optimization

For both mapping-unit schemes, training samples were constructed from the complete set of spatial mapping units rather than using all mapping units directly for model training. Mapping units containing at least one documented landslide point were identified as landslide-positive samples. Non-landslide samples were selected from mapping units without documented landslides and located outside a predefined exclusion distance from known landslide locations, in order to reduce the possibility of selecting potentially unstable areas as negative samples.

For each mapping-unit scheme, the number of non-landslide samples was set equal to the number of landslide-positive samples, resulting in a 1:1 positive-to-negative sample ratio. Therefore, the total numbers of raster cells and slope units differed substantially because of their different spatial representations, whereas the datasets used for model training were constructed according to the same balanced sampling principle.

Accordingly, a positive label indicates the presence of a documented landslide within the evaluation unit rather than instability over the entire spatial extent of that unit. Because the available landslide inventory is point-based rather than polygon-based, the proportion of landslide-affected area within individual mapping units was not used as a labeling criterion. After spatial overlay, 128 raster cells and 112 slope units were identified as landslide-positive samples, yielding final modeling datasets of 256 raster-cell samples and 224 slope-unit samples, respectively.

To determine the optimal model configurations, the samples under each mapping-unit scheme were first randomly divided into a training subset and a held-out test subset. The training subset was used for model fitting and hyperparameter optimization, whereas the held-out test subset was excluded from both model training and hyperparameter selection and was used only for the final model-performance evaluation.

Hyperparameter tuning for LR, MLP, RF, and XGBoost was conducted exclusively within the training subset using a grid-search procedure combined with five-fold cross-validation (Wei et al., 2025). Because the two mapping-unit schemes represent the conditioning factors in different spatial forms and yield different data distributions, the same optimal hyperparameter configuration was not assumed a priori for both schemes. Therefore, hyperparameters were tuned independently for each model–unit combination using the same search space, five-fold cross-validation procedure, and optimization criterion.

More specifically, each candidate hyperparameter combination was evaluated within the training subset, and the mean AUC obtained from five-fold cross-validation was used as the optimization criterion. For each mapping-unit scheme, the hyperparameter combination achieving the highest mean validation AUC was selected as the final configuration. The optimized model was subsequently fitted using the training subset and evaluated on the held-out test subset.

This procedure provides a consistent model-training and evaluation framework for comparing the two mapping-unit schemes under the same sampling strategy. The hyperparameter search ranges are listed in Table 2.

TABLE 2. Hyperparameter search ranges.

After model training, hyperparameter optimization, and performance evaluation, the fitted models were applied to all raster cells or all slope units across the study area to calculate the landslide susceptibility probability of each mapping unit and generate the final susceptibility maps.

4.2 Landslide susceptibility mapping and results

Figure 6 presents the landslide susceptibility maps generated by the four machine-learning models under the raster-cell and slope-unit schemes. Higher susceptibility is mainly distributed in the central–eastern and southern mountainous sectors, particularly in areas associated with deeply incised valleys, whereas lower susceptibility is more prevalent in the northwestern high-altitude and densely vegetated areas. Nevertheless, noticeable differences are observed among the model–unit combinations in the spatial extent, continuity, and local configuration of the susceptibility classes.

Figure 7 further quantifies the area proportions of the five susceptibility classes under the different model–unit combinations. The distributions varied among both models and mapping-unit schemes. For XGBoost, the proportion of the very-low-susceptibility class was similar between the raster-cell and slope-unit schemes (0.536 and 0.547, respectively), whereas the very-high-susceptibility proportion increased from 0.0463 to 0.120. For RF, the moderate-susceptibility class accounted for approximately 0.220 under both schemes, while the proportions of the high and very-high classes increased from 0.0985 to 0.0380 under the raster-cell scheme to 0.131 and 0.0682 under the slope-unit scheme. Similar increases in the combined high- and very-high-susceptibility proportions were also observed for MLP and LR. These results indicate that mapping-unit selection affects not only the spatial representation of susceptibility patterns but also the areal composition of susceptibility classes.

Raster cells and slope units produced distinct spatial representation patterns. Under the raster-cell-based scheme (Figures 6a1–d1), the susceptibility maps exhibit a highly fragmented spatial pattern, with high-, moderate-, and low-susceptibility classes interspersed at local scales. Spatial transitions are relatively abrupt, and clear class discontinuities occur within individual slope areas, resulting in fragmented susceptibility boundaries. This pattern is consistent with the fixed-resolution representation of raster cells, through which local variations in topographic and environmental attributes can be retained in the model inputs and subsequently reflected in the susceptibility outputs. Accordingly, raster-cell-based results tend to exhibit stronger short-range spatial heterogeneity and weaker continuity of geomorphological units.

In contrast, results derived from slope units (Figures 6a2–d2) are generally more spatially continuous and stable. High-susceptibility zones tend to form belt-like or patchy distributions along ridge-valley systems and slope-unit boundaries, producing spatial patterns that are more consistent with natural geomorphological structures. Compared with raster cells, slope units better preserve the internal homogeneity of terrain and environmental conditions within each slope, thereby reducing the influence of local-scale variability on model outputs. As a result, susceptibility boundaries are more closely aligned with actual slope structures, and spatial interpretability is significantly improved. For both mapping-unit schemes, most historical landslide points are concentrated within the high- and very-high-susceptibility classes. Across the four models, these two classes contain 83.6%–95.3% of the historical landslide points under the raster-cell scheme and 78.1%–95.3% under the slope-unit scheme. This indicates that both mapping-unit schemes show a generally good correspondence between predicted susceptibility levels and the observed landslide distribution, although the degree of concentration varies among models.

Regarding model differences, all machine-learning models produced broadly consistent susceptibility patterns under the same mapping-unit scheme, although discrepancies persisted in the spatial delineation of highly susceptible zones and the representation of local spatial patterns. The LR model produces relatively smooth results, with continuous high-susceptibility zones but weaker boundary delineation capability. The MLP model captured more complex nonlinear relationships but still showed local spatial fluctuations. RF and XGBoost generally showed stronger ability to capture spatial structures, producing more concentrated high-susceptibility zones and clearer gradational transitions. Among the evaluated models, XGBoost exhibited the clearest spatial discrimination of highly susceptible areas.

Overall, variations in mapping-unit design markedly alter landslide-susceptibility mapping results, regardless of which predictive algorithm is applied. Raster cells tend to emphasize local spatial heterogeneity and produce more fragmented susceptibility patterns, whereas slope units emphasize the integrity and continuity of geomorphological units, offering superior spatial representation and geomorphic interpretability. Therefore, mapping-unit selection affects not only the structure of model inputs but also the spatial distribution and geological interpretability of landslide susceptibility results.

4.3 Uncertainty analysis of prediction results

4.3.1 Model accuracy evaluation

To further investigate the effect of mapping-unit selection on LSA outcomes, model performance was quantitatively assessed using accuracy, precision, recall, F1 score, and AUC.

As shown in Figure 8, the predictive performance differed systematically between the two mapping-unit schemes. Across the four models, the mean AUC increased from 0.798 under the slope-unit scheme to 0.859 under the raster-cell scheme, corresponding to an increase of 0.061. Accuracy and F1 score showed a similar overall pattern, with generally higher values under the raster-cell scheme. These quantitative results indicate that, under the present validation framework, raster-cell-based models provide stronger discrimination between landslide and non-landslide samples.

Under the raster-cell framework, RF and XGBoost showed the best overall performance. Their AUC values reached 0.927 and 0.913, respectively, compared with 0.835 for MLP and 0.762 for LR. RF achieved the highest accuracy (0.870), precision (0.888), recall (0.869), and F1 score (0.868), while XGBoost also performed strongly, with corresponding values of 0.818, 0.833, 0.817, and 0.816. These results quantitatively demonstrate that the ensemble-learning models outperformed LR and MLP under the raster-cell scheme.

Under the slope-unit scheme, overall performance metrics decreased slightly across all models; however, inter-model performance differences were substantially reduced, indicating a more convergent trend. RF and XGBoost still maintain relatively superior performance, but the gap between them and the other models becomes smaller. This suggests that slope units partially reduce variability caused by model structure, leading to more consistent results.

Comparatively, raster cells achieved higher overall predictive accuracy but showed greater inter-model variability, whereas slope units slightly reduced overall accuracy but improved consistency and stability across models. Therefore, the two mapping unit schemes exhibit a trade-off between prediction accuracy and result stability.

Overall, raster cells are more advantageous for improving predictive accuracy, while slope units contribute to enhancing the consistency and robustness of model outputs.

4.3.2 Analysis of LSI indicators

Here, the overall level of the predicted LSI values was characterized using the mean statistic, while the degree of spread among these predictions was measured using the standard deviation. As shown in Figure 9, the LSI distributions differed significantly between the two mapping-unit schemes and generally showed low-value concentration and high-value dispersion.

In terms of mean LSI, the same model showed clear differences between the two mapping-unit schemes. For the LR model, the mean LSI increases from 0.327 under the raster-cell scheme to 0.368 under the slope-unit scheme. For the MLP model, the mean increases from 0.259 to 0.335, indicating that slope units generally lead to higher LSI values. In contrast, ensemble learning models show an opposite trend: RF decreases from 0.280 (raster) to 0.244 (slope), and XGBoost decreases from 0.245 to 0.192. This suggests that slope units suppress the diffusion of locally high susceptibility values in ensemble models, leading to more conservative predictions. These results indicate that the influence of mapping units on LSI mean values is strongly model-dependent.

In terms of standard deviation, the influence of mapping units on the dispersion of LSI is even more pronounced. Under the raster-cell scheme, MLP and XGBoost exhibit higher standard deviations (0.277 and 0.212, respectively), indicating that raster -based representations tend to amplify local spatial heterogeneity. In contrast, under the slope-unit scheme, LR and RF show significantly lower standard deviations (0.213 and 0.225, respectively), suggesting that geomorphological constraints enhance spatial homogeneity and produce more concentrated LSI distributions. Overall, slope units tend to reduce dispersion, whereas raster cells enhance the expression of spatial heterogeneity.

Overall, mapping unit type has a significant influence on LSI distribution characteristics. Specifically, the direction and magnitude of mapping-unit effects on mean LSI varied among models, indicating that LSI distribution is jointly controlled by mapping-unit type and model structure. In contrast, raster cells tend to produce distributions dominated by lower values, although their dispersion is more sensitive to the underlying model structure.

4.3.3 Reasonableness validation

As a commonly used reasonableness assessment in landslide susceptibility studies, the distribution of historical landslides across different susceptibility classes was examined by calculating the landslide density for each class under the two mapping-unit schemes. This analysis aims to evaluate the spatial consistency and reasonableness of the susceptibility zonation rather than to serve as an independent validation of model predictive performance. The corresponding results are summarized in Table 3.

TABLE 3. Statistical results of landslide susceptibility mapping.

Table 3 shows a common distributional trend across all eight model–unit combinations. Historical landslides are relatively sparse in the very-low- and low-susceptibility classes, whereas landslide density generally increases toward the high- and very-high-susceptibility classes. Notably, the highest landslide density is consistently observed in the very-high-susceptibility class for all model–unit combinations. This progressive concentration of historical landslides toward higher susceptibility classes is consistent with the expected distribution pattern of susceptibility zonation and therefore supports the reasonableness of the resulting susceptibility classification (Zhao et al., 2024).

More specifically, under the raster-cell scheme, the ensemble-learning models RF and XGBoost showed the highest landslide densities in the very-high-susceptibility class, with peak values of 20.989 and 17.872, respectively. These values were substantially higher than those of LR and MLP, indicating the stronger ability of ensemble models to identify high-risk areas. Under the slope-unit scheme, all models also show a stable increasing trend in landslide density with susceptibility level, with the very high class concentrating most landslide occurrences (ranging from 4.327 to 12.721). However, the peak values were lower than those under the raster-cell scheme, suggesting that slope units enhance spatial structural consistency but may weaken local extreme responses.

Compared with slope units, raster cells tended to enhance local high-risk responses and produce higher density peaks, whereas slope units emphasized spatial consistency and landslide clustering at the geomorphological-unit scale.

5 Discussion

5.1 Influence of mapping units on model decision basis

To analyze the influence of mapping-unit type on model decision mechanisms, SHAP-based interpretability analysis was conducted using the RF model, which showed the best overall performance among the four machine-learning models. Therefore, its feature-attribution results can better reveal differences in model decision mechanisms between the two mapping-unit schemes.

As shown in Figure 10, the RF model identified a generally consistent set of dominant conditioning factors under the two mapping-unit schemes. RoadDis, RiverDis, and rainfall exhibited relatively high mean SHAP values, indicating that rainfall input, fluvial erosion, and anthropogenic disturbance play important roles in predicting landslide susceptibility in the study area. In contrast, lithology showed a relatively low SHAP contribution under both schemes. However, this result reflects only the relative importance of lithology within the current RF model and should not be interpreted as evidence of a weak physical control of lithology on landslide development. SHAP values are also jointly influenced by the spatial distribution of samples, the information provided by individual predictors, correlations and interactions among conditioning factors, and the fitted model structure.

In this study, lithology was derived from a 1:200,000 geological map and generalized into four engineering-geological groups; consequently, neighboring cells within the same geological unit share the same lithological category. The available geological data therefore cannot resolve fine-scale lithological variations below the original mapping scale, and the present results should not be used to infer detailed lithological controls at the scale of individual landslides. A rigorous assessment of geological-map scale effects would require independent geological datasets at different mapping scales, particularly finer-scale geological information. Simply resampling the same 1:200,000 lithological dataset to different grid resolutions would not introduce additional geological information and thus would not provide an effective sensitivity test of geological-map scale.

Under the raster-cell scheme (Figure 10a), the model shows higher sensitivity to local environmental heterogeneity. RoadDis and RiverDis ranked among the most influential variables and showed strong contributions to the model predictions, indicating that the raster-cell-based RF model relies strongly on information related to road proximity and river proximity. In addition, relief, aspect, and rainfall also showed relatively high contributions, indicating that raster-based units are more effective in capturing fine-scale topographic-hydrological coupling. In contrast, NDVI and SPI contribute relatively less and exhibit more scattered spatial patterns.

In comparison, under the slope-unit scheme, the model shows greater dependence on geomorphological information. Precipitation remains one of the most influential variables, with its relative contribution further increasing (Figure 10b). Meanwhile, the contributions of Aspect, PlanCurv, and TerrRough are enhanced, indicating greater model dependence on slope geometry and terrain-continuity information. Meanwhile, the contributions of Aspect, PlanCurv, and TerrRough are significantly enhanced, indicating a higher dependence on slope geometry and terrain continuity information. This suggests that slope units reduce the influence of local perturbations by integrating geomorphological information at the slope-unit scale, thereby shifting the model decision basis from locally driven heterogeneity toward structurally constrained mechanisms.

Overall, although the mapping unit type does not alter the fundamental set of dominant conditioning factors, it significantly affects the dependence of the model on different types of environmental information. Raster cells emphasize local disturbance signals and micro-topographic heterogeneity, whereas slope units highlight geomorphological structure and regional environmental constraints.

5.2 Applicability analysis of mapping units

As two typical spatial representation schemes in LSA, raster cells and slope units exhibited distinct advantages in Hejing County. In this study, their applicability was assessed primarily in terms of predictive performance, spatial representation, inter-model consistency, and geomorphological interpretability. The results show that raster cells generally achieved higher predictive performance and were more sensitive to local spatial heterogeneity, whereas slope units produced more spatially continuous susceptibility patterns, reduced inter-model variability, and better preserved geomorphological structure. These differences indicate that the suitability of a mapping-unit scheme depends on the specific objectives and spatial scale of the LSA application.

As shown in Figure 11, raster cells discretize the study area at a fixed spatial resolution and retain local variations in conditioning factors with relatively limited spatial aggregation. This representation allows small-scale differences in topography, vegetation, hydrological conditions, and anthropogenic disturbance to remain distinguishable in the model inputs. Consequently, raster-cell-based models can respond more sensitively to localized environmental contrasts. However, this local sensitivity also increases the likelihood that short-range variations are translated into fragmented susceptibility patches and abrupt class boundaries. Thus, the predictive advantage of raster cells is accompanied by a trade-off in spatial continuity and geomorphological coherence. This characteristic makes raster cells particularly suitable for site-specific susceptibility assessment and infrastructure siting, where localized spatial detail is important, but less advantageous when the objective is to represent the integrity of natural slope systems.

In contrast, slope units aggregate conditioning-factor information within terrain-defined slope boundaries, thereby reducing the influence of highly localized variations and emphasizing the integrated characteristics of individual slopes. This aggregation effect may partly explain why the performance differences among models become smaller under the slope-unit scheme, because model predictions are based on spatially smoothed and geomorphologically structured inputs rather than highly localized cell-level contrasts. Although such aggregation may reduce sensitivity to small-scale susceptibility variations, it improves the continuity of susceptibility patterns and their correspondence with natural slope boundaries. Therefore, slope units are more suitable when the objective is regional-scale susceptibility zonation and geomorphologically coherent spatial interpretation.

The effective spatial scale of a mapping unit is also related to landslide-inventory representation and slope-unit construction. Although the same point-based inventory and sampling strategy were used for both schemes in this study, the 128 landslide records corresponded to 128 positive raster cells but 112 positive slope units, indicating that the same landslide information is represented over different spatial supports. DEM resolution further affects this relationship. The 30 m DEM determines the resolution of terrain attributes in both schemes, but also controls flow routing and ridge–valley identification during slope-unit delineation. In addition, parameters such as the flow-accumulation threshold, minimum unit area, and circular-variance threshold determine the spatial granularity and degree of environmental aggregation within slope units. Therefore, slope units derived using different DEM resolutions and delineation settings may represent different effective geomorphological scales, which can partly explain why their relative performance varies among study areas.

In summary, mapping-unit selection reflects a trade-off between local sensitivity and geomorphological structural consistency in LSA. Raster cells are better suited for capturing fine-scale spatial heterogeneity, whereas slope units are more effective in revealing large-scale geomorphological controls and continuous spatial patterns. Accordingly, slope units should be prioritized for regional-scale susceptibility mapping to obtain stable and geomorphologically meaningful zonation results, whereas raster cells can be used in high-risk or engineering-sensitive areas for fine-scale identification and localized risk characterization. This combined strategy can support a multi-scale framework integrating regional zonation and local refinement.

5.3 Considerations regarding the validation strategy

In this study, the samples were randomly divided into training and test subsets, and five-fold cross-validation was conducted within the training subset for hyperparameter optimization. Random partitioning combined with cross-validation is commonly used for model training and validation in landslide susceptibility assessment. The primary objective of this study is to compare the effects of raster cells and slope units on model performance and susceptibility assessment under a consistent data and modeling framework. Therefore, the same data-partitioning, hyperparameter-optimization, and performance-evaluation procedures were applied to both mapping-unit schemes to maintain consistent comparison conditions.

A major limitation of random partitioning is that it does not explicitly account for spatial autocorrelation among samples. Spatially adjacent raster cells or slope units may share similar topographic, hydrological, geological, and other environmental characteristics, such that geographically close samples with similar attributes may occur in both the training and test subsets. Compared with spatial block cross-validation, random partitioning imposes a weaker constraint on spatial independence between training and test samples and may therefore yield relatively optimistic performance estimates when the objective is to assess model extrapolation to spatially independent areas. This issue deserves particular attention for fine-resolution raster cells because neighboring cells may exhibit relatively strong local spatial dependence.

Spatial block cross-validation can reduce this effect by geographically separating training and validation samples; however, it also has inherent limitations. Its evaluation results can be sensitive to the size, number, location, and boundaries of the spatial blocks. Different blocking schemes may alter the number of training and validation samples, the proportions of landslide and non-landslide samples, and the distributions of environmental covariates. When blocks are excessively large, the reduced training sample size and increased environmental differences between training and validation areas may cause the evaluation to reflect both spatial and environmental extrapolation. Conversely, blocks that are too small may not sufficiently reduce spatial autocorrelation. Moreover, raster cells and slope units differ substantially in spatial scale, geometry, and boundary structure, and a common blocking configuration may not impose equivalent spatial constraints on the two mapping-unit schemes. Consequently, spatial block validation is itself scale-dependent and should not be regarded simply as a direct replacement for random validation.

These considerations indicate that the two validation strategies emphasize different aspects of model performance. Spatial block validation is particularly useful for assessing generalization and extrapolation under spatial separation, whereas the present study focuses on the relative differences between mapping units within the same study area, data basis, and modeling framework. Because identical random-partitioning principles, optimization procedures, and evaluation metrics were applied to the raster-cell- and slope-unit-based models, the adopted validation strategy is adequate for the controlled comparative objective of this study. Accordingly, the results are interpreted primarily as the relative effects of mapping-unit selection under a unified validation framework rather than as direct estimates of model extrapolation performance in spatially independent regions. Future studies could further compare random validation with spatial blocking at multiple spatial scales to quantify the effects of spatial autocorrelation and block configuration on the relative performance of different mapping units.

6 Conclusion

Using Hejing County, Xinjiang, as a case study, this work provides a comprehensive evaluation of how raster cells and slope units influence the performance of LSA, spatial representation, and model interpretability. The main conclusions drawn from this study are outlined below.

• Raster cells improved overall predictive accuracy, whereas slope units enhanced the consistency and stability of predictions across models. Under the raster-cell scheme, all models achieve higher performance metrics, indicating better predictive capability. In contrast, under the slope-unit scheme, inter-model performance differences were substantially reduced, and the results became more consistent, demonstrating stronger robustness in multi-model comparison.
• Raster cells exhibit unique advantages in capturing local heterogeneity and identifying high-risk patches. They are more sensitive to local variations in conditioning factors and can effectively delineate small-scale high-susceptibility zones. The higher standard deviation of LSI under the raster-based scheme indicates stronger contrast between high- and low-susceptibility areas, making it suitable for fine-scale hazard assessment and engineering applications.
• The choice of mapping unit significantly influences model decision logic and factor contribution structure. SHAP analysis indicates that slope units emphasize a “geomorphological structure–hydrological process” control framework, whereas raster cells place greater emphasis on local heterogeneity factors. These findings suggest that mapping-unit selection should be guided by the objective of the analysis: slope units are recommended for regional-scale hazard zonation, whereas raster cells can be used for supplementary fine-scale analysis in local high-risk areas.

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.

Footnotes

Edited by: Srđan Kostić, Jaroslav Černi Water Institute, Serbia

Reviewed by: Zhen Zhang, Anhui University of Science and Technology, China

Yong Chen, China Three Gorges University, China