Landslide Susceptibility Assessment in Ya’an Based on Coupling of GWR and TabNet

Abstract

Landslides are destructive geological hazards, making accurate landslide susceptibility assessment essential for disaster prevention and mitigation. However, existing studies often lack scientific rigor in negative sample construction and have unclear model applicability. This study focuses on Ya’an City, Sichuan Province, China, and proposes an innovative approach to negative sample construction using Geographically Weighted Regression (GWR), which is then integrated with Tabular Network (TabNet), a deep learning architecture tailored to structured tabular data, to assess landslide susceptibility. The performance of TabNet is compared against Random Forest, Light Gradient Boosting Machine, deep neural networks, and Residual Networks. The experimental results indicate that (1) the GWR-based sampling strategy substantially improves model performance across all tested models; (2) TabNet trained using the GWR-based negative samples achieves superior performance over all other evaluated models, with an average AUC of 0.9828, exhibiting both high accuracy and interpretability; and (3) elevation, land cover, and annual Normalized Difference Vegetation Index are identified as dominant predictors through TabNet’s feature importance analysis. The results demonstrate that combining GWR and TabNet substantially enhances landslide susceptibility modeling by improving both accuracy and interpretability, establishing a more scientifically grounded approach to negative sample construction, and providing an interpretable, high-performing modeling framework for geological hazard risk assessment.

1. Introduction

Landslides are spatial displacement processes involving the downslope movement of rocks, soil, and loose debris, either as a coherent mass or in scattered form, driven by tectonic activity, hydrometeorological factors, and gravitational forces [1]. In China, landslides primarily occur in hilly and mountainous areas and are characterized by wide distribution, delayed onset, and significant destructiveness. Accurate landslide susceptibility assessment is vital for early warning and disaster mitigation efforts. Ya’an City, situated in the transitional zone between the Sichuan Basin and the Tibetan Plateau in southwestern China, features complex topography, rugged terrain, and active tectonic activity, resulting in a high frequency of landslide occurrences. Assessing landslide susceptibility in Ya’an offers both strong regional representativeness and substantial practical relevance.

Current landslide susceptibility assessment methods can be broadly categorized into two groups: qualitative approaches based on expert judgment and quantitative approaches grounded in geostatistics and machine learning [2,3]. Qualitative methods such as hierarchical analysis [4] and fuzzy comprehensive evaluation [5] rely heavily on expert experience and are often unsuitable for large-scale or objective assessments. Statistical models, including the information value model [6], the weight of evidence model [7], and the frequency ratio model [8], use mathematical formulas to approximate variable relationships. However, these methods often struggle to capture nonlinear and complex interactions among environmental factors. In contrast, machine learning approaches, such as Boosted Regression Trees (BRTs) [9], Support Vector Machines (SVMs) [10,11], Random Forests (RFs) [12], and Light Gradient Boosting Machine (LightGBMs) [13], have been proven to be able to capture the complex interactions and demonstrate good accuracy. However, their performance may deteriorate on large-scale datasets or in the presence of complex feature interactions due to limitations in representational capacity and generalization ability [14].

The emergence of deep learning has introduced models such as deep neural networks (DNNs) [15], Convolutional Neural Networks (CNNs) [16], Recurrent Neural Networks (RNNs) [17], and Long Short-Term Memory (LSTM) [18] networks into landslide susceptibility modeling. More advanced architectures such as Residual Networks (ResNet) [19] and U-Net [20] further enhance the ability to capture complex spatial patterns and extract high-level features. However, due to differences in architecture and learning mechanisms, these models exhibit varying strengths and suitability for different tasks: DNNs are suited for structured tabular data, CNNs excel at extracting spatial features from images, RNNs and LSTMs are designed for temporal sequence modeling, and ResNet/U-Net architectures are effective in learning deep representations and spatial hierarchies. In landslide susceptibility assessment, which is primarily driven by multi-factor structured datasets, not a single deep learning model is universally optimal. Deep learning approaches face limitations in training efficiency, interpretability, and their adaptability to tabular data formats.

To overcome these limitations, this study introduces Tabular Network (TabNet)—a deep learning architecture specifically developed for structured data. Proposed by Google Cloud in 2019 [21], TabNet has gained popularity in domains such as financial risk assessment [22] and medical diagnosis [23], and has also been applied in environmental [24] and remote sensing research [25]. However, its application to landslide susceptibility analysis remains largely unexplored. By integrating the interpretability and sparse feature selection mechanisms of tree-based models with the end-to-end learning capabilities of neural networks, TabNet offers a promising framework for spatial risk modeling. To assess its performance, we compare TabNet against traditional machine learning models (RF, LightGBM) and deep learning models (DNN, ResNet).

One of the fundamental principles of machine learning is that training datasets must include both positive samples and negative samples, whose quality plays a critical role in determining model performance. In landslide susceptibility modeling, positive samples refer to locations where landslides have already occurred, typically derived from historical landslide inventories. Negative samples refer to locations where no landslide has been recorded or where the terrain is currently considered stable without visible signs of slope failure, and are usually selected from areas where landslides have not occurred. However, randomly selected non-landslide areas may contain areas that may be in the early stages of instability, thereby introducing label noise and adversely affecting the predictive performance of the model [26]. Most existing studies rely on random sampling, which does not guarantee the reliability of negative samples and may result in the mislabeling of high-risk areas as negatives, thereby undermining predictive accuracy.

Geographically Weighted Regression (GWR), a spatial regression technique that accounts for spatial nonstationarity, produces local coefficients that capture the spatial variability in factor contributions. Coefficients with low absolute values indicate weak explanatory power and limited linear correlation with landslide occurrence [27]. GWR has been widely applied in spatial partitioning and hazard sensitivity assessment, demonstrating strong interpretability in spatial risk evaluation [28]. Although some studies have attempted to incorporate GWR into negative sample selection [29], such applications remain relatively uncommon and lack a standardized or systematic framework. In this study, we compute the GWR coefficients for each landslide susceptibility indicator, apply absolute value transformation and normalization, and subsequently aggregate the coefficients to derive a composite explanatory index. Negative samples are selected from regions with low composite scores to reduce the likelihood of mislabeling potentially high-risk areas, thus improving the distinguishability and reliability of training data.

In summary, this study focuses on Ya’an City in Sichuan Province, China, and integrates GWR with TabNet to assess landslide susceptibility. We conduct a comparative analysis of GWR-based and traditional random negative sampling strategies, and evaluate four modeling approaches (RF, LightGBM, DNN, ResNet) alongside TabNet to examine how different negative sampling strategies and modeling methods influence the uncertainty of landslide susceptibility predictions. Through a novel sample construction strategy and tailored model architecture, the proposed approach significantly enhances predictive accuracy and spatial adaptability, offering both technical support and methodological guidance for landslide risk management.

2. Study Area and Data Sources

2.1. Study Area

Ya’an City is located in the central-western part of Sichuan Province, China, spanning from 28°51′ to 30°56′N latitude and 101°56′ to 103°23′E longitude, as shown in Figure 1. Ya’an is located in the transitional zone between the Tibetan Plateau and the Sichuan Basin, characterized by significant elevation differences and complex terrain which define its distinctive topography and geomorphology. Situated along the boundary between the Indian and Eurasian plates, the region features dense fault systems and diverse lithological units, creating favorable conditions for landslides, a result of its geological structure. Its subtropical humid monsoon climate brings abundant, concentrated rainfall, especially during the rainy season, often triggering slope failures through surface runoff accumulation. In addition, human activities such as road construction and urban development also disturb natural slopes, further increasing the risk of landslides. Therefore, selecting Ya’an as the study area is both representative and provides valuable insights for landslide susceptibility assessment in regions with similar geomorphological and climatic characteristics.

2.2. Data Sources

The landslide dataset was constructed by integrating multiple sources, including results from geological hazard investigations and landslide point data released by local governments or relevant agencies, resulting in the identification of 1481 historical landslide points in Ya’an City. All landslide points were standardized to the same coordinate reference system and checked for alignment with environmental raster layers. Spatial outliers and duplicate entries were removed through visual inspection to ensure spatial consistency and reliability.

Landslides are triggered by a combination of natural and anthropogenic factors, involving topographic, geological, hydrometeorological, and human activity-related elements. To systematically capture these potential influencing factors, this study draws extensively on previous research and incorporates the geological characteristics and landslide development context of the study area. A total of 17 indicators were initially selected for the landslide susceptibility assessment, including the following: elevation, slope, aspect, Slope Length and Steepness Factor (LS) [30], Terrain Ruggedness Index (TRI) [31], plan curvature [32], profile curvature [33], Topographic Wetness Index (TWI) [34], Stream Power Index (SPI) [35], lithology, distance to faults, earthquake density, annual precipitation, annual Normalized Difference Vegetation Index (NDVI) [36], distance to rivers, land cover, and distance to roads.

The data sources corresponding to these indicators are summarized in

Table 1. Data sources.

Data Type	Data Name	Resolution	Data Source
Historical Landslide Point Data	Spatial Distribution Data of Geological Hazard Points in China	-	Resource and Environmental Science Data Platform
Elevation Data	ASTER GDEM Data of Ya’an City	30 m	Geospatial Data Cloud site
LS Data	Slope Length and Steepness Factor Dataset for the 20 Countries of the Pan-Third Pole Region	7.5 arc-s	National Tibetan Plateau Data Center
Lithology Data	Spatial Lithology Distribution Data of China	7.5 arc-s	Resource and Environmental Science Data Platform
Fault Vector Data	Vector Data of Active Faults in China	-	Seismic Active Fault Survey Data Center
Earthquake Epicenter Data	Global Earthquake Data (2000–2022)	-	United States Geological Survey
Annual Precipitation Data	Annual Precipitation Data of China (2000–2022)	1 km	National Earth System Science Data Center
River and Road Vector Data	Fundamental Geographic Information Data of China	-	National Geomatics Center of China
Annual NDVI Data	Annual Average NDVI Data of China (2000–2022)	1 km	Resource and Environmental Science Data Platform
Land Cover Data	Land Cover Data of China	30 m	National Geomatics Center of China

. For indicators without explicit citations, the data were obtained from existing datasets and processed using GIS-based techniques. According to existing research, utilizing a grid unit with a resolution of 30 m or finer in landslide susceptibility analysis enhances the model’s predictive capability [37,38]. Therefore, all indicators were standardized to the same spatial resolution of 30 m and temporal scale, and were uniformly preprocessed. The final processed datasets are presented in Figure 2.

2.3. Division of Evaluation Units and Selection of Susceptibility Indicators

This study uses slope units as the fundamental spatial unit for landslide susceptibility assessment. Slope units are delineated based on valley boundaries and are widely recognized for their effectiveness in capturing local terrain and geomorphological features [39]. They also exhibit strong spatial control over the development of geological hazards such as landslides, debris flows, and rockfalls, making them a commonly used unit in hazard susceptibility evaluations [40]. In this study, slope units were generated using hydrological analysis tools in the GIS platform. The main processing steps included sink filling, flow direction extraction, flow accumulation calculation, stream network generation, and watershed segmentation [41]. A total of 16,491 slope units were delineated across the study area.

As a preliminary step, this study uses the Geodetector method to filter out those susceptibility indicators that have low relevance to landslide occurrence [42]. Geodetector is a statistical tool based on the principle of spatial stratified heterogeneity [43]. Its core metric, the q-value, quantifies the extent to which a given factor explains the spatial distribution of a target variable. A q-value closer to 1 indicates stronger explanatory power. As shown in Figure 3a, the q-value results were calculated, and a threshold of 0.05 was applied in this study [44,45]. Seven indicators with weak spatial explanatory power were excluded, and the remaining ten were retained for subsequent modeling.

To reduce potential interference from multicollinearity among indicators, this study performed a correlation analysis. Pearson correlation coefficients [46] were calculated between all pairs of susceptibility indicators to quantitatively assess the degree of correlation. As shown in Figure 3b, there is a strong correlation between TRI and slope, as well as between annual precipitation and elevation, with coefficients exceeding 0.7 [47]. Due to this strong correlation, TRI and annual precipitation were excluded, and the remaining eight indicators were retained as input variables for the modeling phase.

3. Method

The overall workflow of this study is illustrated in Figure 4 and consists of four main stages:

Data preparation and preprocessing: This step includes data collection, the division of evaluation units, and the selection of susceptibility indicators (described in Section 2.2 and Section 2.3).

Sample data construction: Historical landslide points serve as positive landslide samples, while negative landslide samples are generated using the GWR-based negative sample construction method. A random sampling-based negative sample construction is also included for comparison.

Model training phase: Given the generated samples from the previous step, TabNet is trained and compared with several widely used methods, including RF, LightGBM, DNN, and ResNet. Model performance is assessed using several commonly used metrics.

Prediction phase: The best-performing TabNet model is used for landslide susceptibility prediction, followed by feature importance analysis, offering support for disaster prevention, risk management, and spatial planning.

3.1. Negative Sample Construction Based on GWR

This study introduces the GWR model to guide the selection of negative samples. The workflow of GWR-based negative sample construction process is illustrated in Figure 5a.

3.1.1. GWR Modeling Principle

GWR is a local regression technique that accounts for spatial nonstationarity by fitting a separate regression model at each spatial location [27], which enables a more precise representation of spatial heterogeneity. The local regression coefficients produced by GWR represent the spatial variation in the explanatory power of individual factors. In this study, GWR is applied using the eight selected susceptibility indicators from Section 2.3 as independent variables, and the number of historical landslide occurrences within each slope unit as the dependent variable. The general form of the GWR model is expressed as follows:

$$y_i={\beta }_0\left(u_i,v_i\right)+{\sum }_k{\beta }_k\left(u_i,v_i\right)x_{ik}+{\epsilon }_i.$$

In this equation: $y_i$ is the number of landslide occurrences in the $i^{th}$ slope unit; $x_{ik}$ is the value of the $k^{th}$ susceptibility indicators for the $i^{th}$ slope unit; $\left(u_i,v_i\right)$ denotes the geographic coordinates of the $i^{th}$ unit; ${\beta }_k\left(u_i,v_i\right)$ is the regression coefficient of the $k^{th}$ indicators at location $\left(u_i,v_i\right)$; and ${\epsilon }_i$ is the random error term.

3.1.2. Identification of Low-Explanatory Areas and Selection of Negative Samples

The GWR model reveals spatial nonstationarity in the relationships between independent and dependent variables, with its regression coefficients reflecting spatial variability in the explanatory capacity of each factor with respect to landslide occurrence. Note that the purpose of this process is not to directly assess landslide susceptibility, but rather to identify areas where the model exhibits weak explanatory performance. When the absolute values of regression coefficients in a given area approach zero, it indicates that the corresponding factors have limited or unstable contributions to landslide occurrence in that region. These areas can be regarded as low-explanatory zones, where the model struggles to distinguish between landslide-prone and stable units, making them suitable candidates for negative sample selection.

To construct reliable negative samples, this study computed the absolute values of the GWR coefficients for the eight selected susceptibility indicators. These values were then normalized and aggregated to produce a composite explanatory index, where lower index values indicate weaker overall explanatory capacity for landslide occurrence. The natural breaks classification method [48]—a data clustering technique that minimizes within-class variance and maximizes between-class variance for optimal classification—was applied to divide the index into five levels, as it is widely used in geoscientific research, particularly in spatial risk zoning [49,50]. And the two lowest levels, which represent regions with the weakest explanatory power, were selected as low-explanatory zones. Within these zones, negative samples equal in number to the positive samples were randomly selected. This approach helps avoid the mislabeling of potentially high-risk areas as negative samples, thereby improving the reliability of the training data and enhancing the generalization performance of the model.

To evaluate the effectiveness of the GWR-based negative sample construction method, a baseline approach based on conventional random sampling was also implemented. Specifically, a 100 m buffer [49,51] was generated around each historical landslide point, and an equal number of negative samples were randomly selected from areas outside these buffer zones.

3.2. Landslide Susceptibility Modeling Based on TabNet

This study employs TabNet to model landslide susceptibility. To comprehensively evaluate modeling performance, several widely used methods are selected for comparison, including traditional machine learning models (RF and LightGBM), as well as deep learning models (DNN and ResNet). The overall modeling workflow is illustrated in Figure 5b.

3.2.1. Overview of TabNet

The architecture of TabNet is illustrated in Figure 6. TabNet preserves the end-to-end and representation learning characteristics of deep neural networks, while also incorporating the interpretability and sparse feature selection of tree-based models. This hybrid design enables TabNet to deliver strong predictive performance on tabular data tasks, often matching or exceeding that of mainstream tree-based algorithms, while also providing enhanced model transparency. Given that landslide susceptibility assessment is a task dominated by structured and multi-factor data, TabNet demonstrates strong adaptability to this setting.

TabNet consists of multiple decision steps, each employing feature and attentive transformers for dynamic feature selection and nonlinear transformation. The input features are first processed through Batch Normalization. At each decision step, the Attentive Transformer generates a sparse feature mask that dynamically selects the most relevant subset of features. This mask is applied to the input before passing it to the Feature Transformer, which consists of multiple fully connected layers with ReLU activations to enhance the model’s nonlinear representation capacity. The transformed output is then divided by the split module into two branches: one is aggregated across all steps via the aggregation module, while the other is propagated to the next decision step for further feature selection and transformation. Finally, the accumulated output from all decision steps is fed into a fully connected layer to generate the final prediction.

During model training, the TabNet model is optimized using the Adam optimizer, with binary cross-entropy as the loss function. All input features are standardized before training. The core hyperparameters of TabNet are configured as follows: both the feature transformer and decision step dimensions are set to 64; the number of decision steps is 3; the relaxation factor is 1.2; and the sparsity regularization coefficient is set to 0.01. All other parameters are kept at their default values. The model is trained for a maximum of 100 epochs with a batch size of 256.

To assess the applicability and effectiveness of the TabNet model, this study introduces four widely used comparative models for training and evaluation. These models include RF [52], which constructs multiple decision trees by bootstrapping the training data and aggregates their predictions through majority voting; LightGBM [53], an efficient implementation of the Gradient Boosting Decision Tree framework that employs a leaf-wise growth strategy and histogram-based algorithm to accelerate training and improve predictive performance; DNN [54], a feedforward architecture composed of multiple fully connected layers capable of capturing complex nonlinear relationships; and ResNet [55], which extends the DNN structure by incorporating residual connections to mitigate the vanishing gradient problem and enhance training stability.

3.2.2. Cross-Validation

To enhance the stability and generalization of model evaluation, this study adopts a cross-validation approach for training and testing each model. The fundamental idea is to partition the dataset into several subsets, using one subset as the test set and the remaining subsets for training in each iteration [56,57]. This training–validation cycle is repeated multiple times, and the evaluation results are averaged. This method effectively mitigates performance fluctuations caused by the randomness of data partitioning, thereby improving the reliability and representativeness of the evaluation outcomes. Cross-validation has been widely adopted in landslide susceptibility modeling and has been proven to provide a more objective assessment of model performance [58,59].

In this study, a 10-fold cross-validation strategy is employed, in which the dataset is evenly divided into 10 subsets. In each round, one fold is used as the test set and the remaining nine are used as the training set. The average performance across the 10 folds is reported as the final evaluation metric of the model.

3.2.3. Model Performance Evaluation Metrics

This study employs accuracy (ACC), recall (RE), and Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) to evaluate the performance of different models in landslide susceptibility assessment. In addition, ROC curves are used to visualize the models’ discriminative capabilities. To comprehensively assess the agreement between predicted and actual outcomes, as well as the reliability of probabilistic outputs, this study further introduces Cohen’s Kappa coefficient (Kappa) [60] and the Brier score [61] as supplementary evaluation metrics. The definitions and corresponding mathematical formulations of these metrics are summarized in

Table 2. Formulas and definitions of evaluation metrics.

Metric	Formula *	Definition
ACC	$ACC={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$	Proportion of correctly identified landslide and non-landslide units.
RE	$RE={\displaystyle \frac{TP}{TP+FN}}$	Proportion of actual landslide units that are correctly classified as landslides.
ROC	Plot of TPR vs. FPR	Reflects the model’s ability to distinguish landslide from non-landslide areas; the closer the curve is to the top-left corner, the better the performance.
AUC	Area under the ROC curve	Quantifies the model’s discrimination ability; values closer to 1 indicate stronger performance.
Kappa	$Kappa={\displaystyle \frac{p_0-p_e}{1-p_e}}$	Measures agreement between the true classes and the classifications; values closer to 1 indicate stronger consistency, while values near 0 or negative suggest performance equivalent to random guessing, or indicate systematic bias.
Brier Score	$Brier Score={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{(f_i-o_i)}^2$	Measures the mean squared difference between predicted probabilities and actual outcomes; lower scores reflect more reliable probabilistic predictions [62].

* TP (True Positive) and TN (True Negative) represent correctly predicted landslides and non-landslides, respectively, while FP (False Positive) and FN (False Negative) denote misclassified non-landslides and landslides. In the Kappa coefficient, $p_0$ denotes the observed agreement between predicted and actual labels, and $p_e$ denotes the expected agreement by chance. For the Brier Score, $N$ is the total number of samples, $f_i$ is the predicted probability of landslide occurrence for sample $i$, and $o_i$ is the actual observed outcome.

.

4. Results

4.1. Comparison of Results Under Different Negative Sample Construction Strategies

In this study, the eight selected susceptibility indicators are used as independent variables, and the number of landslide occurrences within each slope unit is taken as the dependent variable for GWR fitting. Spatial maps of the regression coefficients for each indicator are shown in Figure 7a–h, illustrating the spatial heterogeneity in their explanatory influence with respect to landslide occurrence. A larger absolute value of a coefficient indicates stronger explanatory influence in that region, while values close to zero suggest a lack of statistically significant association. Positive coefficients denote a positive correlation between the factor and landslide occurrence, whereas negative coefficients indicate an inverse relationship.

Subsequently, the absolute values of the GWR coefficients for all susceptibility indicators were calculated, normalized, and aggregated to a composite explanatory index for each slope unit. This index was then classified into five levels using the natural breaks method, as illustrated in Figure 7i.

A clear spatial correspondence is observed between higher values of the composite explanatory index and the concentration of historical landslide points, validating the spatial rationality of the GWR fitting results. The two lowest index levels, which represent areas with weak overall explanatory power, were selected as candidate zones for negative sample construction. The landslide points in these areas are scattered and very sparse, accounting for only 5% of the total landslides in the study area. Within these zones, an equal number of negative samples were randomly selected to match the number of positive samples. The spatial distribution of the GWR-based negative samples is shown in Figure 8a. For comparison, a conventional random sampling strategy was implemented as a baseline, and the distribution of these randomly selected negative samples is shown in Figure 8b.

Both sets of negative samples were incorporated into the same landslide susceptibility modeling approach, and the resulting AUC values are presented in

Table 3. AUC values * of models based on two negative sample selection methods.

	Negative Samples Selected Based on Random Sampling	Negative Samples Selected Based on GWR
RF	0.9300	0.9661
LightGBM	0.9333	0.9722
DNN	0.9022	0.9632
ResNet	0.9029	0.9618
TabNet	0.9456	0.9828

* Results are presented as the average of 10-fold cross-validation.

. The results demonstrate that the GWR-based negative sampling method substantially enhances predictive performance across all models.

4.2. Comparison of Landslide Susceptibility Modeling Results Across Different Models

The performance metrics for each model are summarized in

Table 4. Performance metrics * of different models.

	ACC	RE	AUC	Kappa	Brier Score
RF	0.9281	0.9489	0.9661	0.8558	0.0601
LightGBM	0.9207	0.9195	0.9722	0.8409	0.0614
DNN	0.9046	0.8974	0.9632	0.8083	0.0720
ResNet	0.8992	0.8903	0.9618	0.7977	0.0735
TabNet	0.9298	0.9196	0.9828	0.8589	0.0521

* Results are presented as the average of 10-fold cross-validation.

and Figure 9. The results demonstrate that TabNet consistently outperforms all other models across most metrics, achieving the highest AUC (0.9828), ACC (0.9298), and the lowest Brier Score (0.0521), indicating both superior discriminative ability and reliable probabilistic prediction. Moreover, TabNet also records the highest Kappa value (0.8589), reflecting a high level of agreement between predicted and actual classifications after adjusting for chance.

Among the traditional machine learning models, RF shows strong performance, achieving the highest recall (0.9489), slightly surpassing TabNet in identifying landslide occurrences. LightGBM also performs competitively, achieving the second-highest AUC (0.9722), though with slightly lower recall and Kappa.

In contrast, the deep learning models DNN and ResNet exhibit comparatively weaker performance. DNN achieves a lower ACC (0.9046) and recall (0.8974), along with a higher Brier Score (0.0720), indicating poorer probability calibration. Despite the use of residual connections, ResNet records a lower ACC (0.8992), recall (0.8903), and Kappa (0.7977), along with the highest Brier Score (0.0735), underscoring its limitations in modeling small to medium-sized structured datasets.

In summary, all evaluated models exhibit distinct levels of predictive capability in landslide susceptibility modeling. Traditional machine learning methods offer stable and interpretable results, making them well-suited for practical applications that require fast and reliable deployment. Deep neural networks demonstrate potential in capturing complex feature dependencies but remain sensitive to data scale and training strategies. TabNet, by contrast, achieves an effective balance among predictive accuracy, interpretability, and adaptability.

4.3. Landslide Susceptibility Prediction and Feature Importance Analysis Based on TabNet

4.3.1. Landslide Susceptibility Prediction

Based on its superior predictive performance, we use TabNet to generate landslide susceptibility predictions across the study area. A continuous susceptibility index ranging from 0 to 1 is produced and subsequently classified into five levels using the natural breaks method. The statistical distribution of these susceptibility classes is summarized in

Table 5. Statistical results of landslide susceptibility zonation using TabNet.

Susceptibility Level	Landslide Count	Landslide Proportion (%)	Area (km2)	Area Proportion (%)	Landslide Density (Events/km2)
Very Low	22	1.49	6716.32	45.00	0.0033
Low	80	5.40	2137.48	14.32	0.0374
Moderate	136	9.18	1481.74	9.93	0.0918
High	156	10.53	1082.48	7.25	0.1441
Very High	1087	73.40	3506.67	23.50	0.3100

and Figure 10.

The results indicate that both the proportion of landslide occurrences and the density of landslide points increase with higher susceptibility levels. The very-high-susceptibility zone, which accounts for only 23.50% of the total study area, contains 73.40% of all recorded landslide events, with a density of 0.3100 events per km², substantially exceeding that of other classes.

The high- and moderate-susceptibility zones contain 10.53% and 9.18% of the total landslide occurrences, with corresponding densities of 0.1441 events per km² and 0.0918 events per km², indicating a moderate degree of risk concentration in these zones.

The low- and very-low-susceptibility zones exhibit substantially lower landslide densities, at 0.0374 and 0.0033 events per km². These zones cover 14.32% and 45.00% of the study area and are considered relatively safe regions with limited landslide activity. The clear spatial delineation of risk levels in the TabNet-generated susceptibility map further validates the model’s reliability and underscores its practical utility in landslide risk management and regional planning.

As shown in Figure 11, the spatial distribution map reveals pronounced spatial heterogeneity in landslide susceptibility across the study area. Very-high-susceptibility zones are primarily concentrated in the eastern, northeastern, and southeastern regions of Ya’an City, where rugged terrain, steep slopes, highly fractured rock masses, and dense fault structures converge. These areas constitute key zones for landslide disaster prevention and control.

High- and moderate-susceptibility zones form patchy and banded patterns surrounding the very high susceptibility regions. They are typically located in areas with abrupt terrain transitions or sharp geological structural changes, reflecting both strong spatial clustering and transitional characteristics of landslide occurrence.

In contrast, low- and very-low-susceptibility zones are predominantly distributed in the northern, northwestern, central, and southern parts of the study area. These areas are characterized by relatively flat terrain, gentle slopes, high vegetation cover, and the presence of concentrated agricultural or built-up land, indicating relatively low landslide risk.

To examine the spatial clustering of landslide occurrences across different landslide susceptibility zones, we performed a hotspot analysis using the Getis-Ord Gi* [63] statistic on both the landslide points and susceptibility classification results, followed by a quantitative interpretation of the outputs. The results are presented in

Table 6. Statistics of hot- and cold spots based on Gi Bin confidence levels.

Gi Bin Value *	p-Value	Meaning	Count	Proportion (%)
3	p < 0.01	99% Confidence Hotspot	349	23.57
2	0.01 ≤ p < 0.05	95% Confidence Hotspot	376	25.39
1	0.05 ≤ p < 0.10	90% Confidence Hotspot	87	5.87
0	p ≥ 0.10	Not Significant	428	28.9
−1	0.05 ≤ p < 0.10	90% Confidence Cold Spot	12	0.81
−2	0.01 ≤ p < 0.05	95% Confidence Cold Spot	41	2.77
−3	p < 0.01	99% Confidence Cold Spot	188	12.69

* The Gi Bin value is a categorical representation of the statistical significance of spatial clustering derived from the Getis-Ord Gi* statistic [64].

and

Table 7. Distribution of hot- and cold spots across landslide susceptibility levels.

Susceptibility Level	Landslide Count	Hotspot Count	Hotspot Proportion (%)	Cold Spot Count	Cold Spot Proportion (%)
Very Low	22	0	0	20	90.91
Low	80	0	0	75	93.75
Moderate	136	5	3.68	102	75
High	156	16	10.26	26	16.67
Very High	1087	791	72.77	18	1.66

and Figure 12. In this context, a hotspot refers to a location where high landslide susceptibility zoning values exhibit statistically significant spatial clustering. Conversely, a cold spot indicates a spatial cluster of low-susceptibility zoning values.

A total of 812 landslide points were identified as hotspots (Gi Bi ≥ 1, p-value < 0.1), representing 54.83% of all landslide points, with 23.57% classified as highly significant hotspots (p-value < 0.01), indicating statistically significant spatial clustering and confirming the non-random nature of landslide distribution. Among these hotspot points, 791 fell within extremely-high-susceptibility zones, accounting for 97.41% of all hotspots and 72.77% of the landslide points in that zone. This demonstrates a pronounced spatial aggregation of landslides in high-susceptibility areas, providing indirect validation for the rationality of the susceptibility zoning scheme. In contrast, cold spots (Gi Bi ≤ −1, p-value < 0.1) were mainly distributed in medium-, low-, and very low-susceptibility zones, comprising less than 20% of all landslide points.

4.3.2. Feature Importance Analysis

To further examine the influence of individual factors on landslide susceptibility, the feature importance scores produced by TabNet model are used to quantitatively evaluate the spatial contribution of each variable, as illustrated in Figure 13.

The results indicate that elevation has the highest importance score at 0.3583, far exceeding all other factors. This suggests that topographic elevation is the dominant driver of landslide occurrence. High-elevation areas are often associated with complex terrain and steep and unstable slopes, and are frequently identified as core zones for landslide activity.

Land cover ranks second in feature importance, with a score of 0.1692, highlighting the significant influence of human activities on slope stability. Construction and agricultural land use frequently involve excavation, terrain modification, and vegetation clearance, all of which may increase the risk of slope disturbance and landslides.

The annual NDVI contributes an importance score of 0.1479, underscoring the protective role of vegetation cover. Areas with high NDVI values typically exhibit dense vegetation, which enhances slope stability by reinforcing soil structure and mitigating erosion caused by rainfall and surface runoff.

Slope (0.0821) and distance to roads (0.0792) demonstrate moderate influence. The former reflects the gravitational component of landslide triggering, while the latter suggests the destabilizing effects of road cuts, drainage alterations, and vibrations from traffic.

Distance to rivers (0.0631) and lithology (0.0562) exhibit lower but non-negligible contributions. River proximity may promote toe erosion and saturation, while lithological properties affect soil strength and water permeability.

Finally, the LS exhibits the lowest importance score at 0.044, indicating a relatively limited role in landslide prediction within the context of this study.

5. Discussion

5.1. Rationale and Limitations of GWR-Based Negative Sample Construction

The results of this study demonstrate that the negative sample construction method based on GWR outperforms the traditional random sampling method across all models. This finding aligns with previous research, suggesting that incorporating spatial distribution characteristics into the negative sample construction process significantly enhances the model’s discriminative ability and generalization performance in landslide susceptibility modeling [65]. This is because traditional random sampling typically selects negative samples from areas without recorded landslides, which may inadvertently include locations that have not experienced landslides but possess high susceptibility. Such incorrect labeling of negative samples can mislead the model and compromise its predictive performance.

Several studies have looked into enhancing negative sample construction methods in landslide susceptibility evaluation, including random selection in areas with slopes of less than 5° [66], random selection outside the landslide buffer zone [67], and selection based on information theory methods [68]. The novelty of our GWR-based negative sample construction method lies in its incorporation of spatial regression information between landslides and influencing factors. This allows the selected negative samples to carry a clearer “low-risk” semantic, thereby enhancing their scientific validity and representativeness. It enables the model to learn more robust discriminative features, providing more accurate and reliable methodological support for landslide susceptibility assessment.

However, the GWR-based negative sample construction method also has certain limitations, particularly its relatively complex computational process. Specifically, its application to large-scale datasets may result in high computational costs due to the need to compute pairwise spatial distances and fit a separate weighted least squares regression at each spatial location. Therefore, improving the computational efficiency of the GWR-based negative sample construction method to meet the demands of large-scale and real-time landslide prediction is a key area for future research. Additionally, considering issues such as data imbalance, spatial heterogeneity, and the complex interactions of influencing factors in landslide susceptibility modeling, future studies should explore diversified negative sample selection methods. These methods should be tailored to the specific characteristics of different datasets and model requirements to further optimize the quality and selection strategy of negative samples.

5.2. Evaluating TabNet’s Robustness and Applicability in Susceptibility Modeling

In terms of model comparison, the TabNet model, a deep neural network specifically designed for structured data, demonstrates exceptional performance, achieving an AUC of 0.9828 and significantly outperforming the other comparison models. This success can be attributed to the unique structural advantages of TabNet. It retains the interpretability and sparse feature selection mechanisms of tree-based models, while also incorporating the end-to-end learning capabilities of neural networks. These combined strengths allow TabNet to more accurately uncover nonlinear relationships and complex features in the data, thereby enhancing both predictive accuracy and model interpretability.

To further evaluate the model’s robustness, we conducted a sensitivity analysis for the TabNet model to evaluate its robustness to key hyperparameters. We systematically varied several important parameters within recommended ranges, as shown in

Table 8. TabNet hyperparameter sensitivity analysis and optimal settings.

Hyperparameter	Description	Tested Range	Optimal Value
n_d	Dimension of the decision layer	[16, 32, 64]	64
n_a	Dimension of the attentive transformer layer	[16, 32, 64]	64
n_steps	Number of decision steps	[3, 5, 7]	3
gamma	Feature reuse penalty coefficient	1.0, 1.2, 1.5	1.2
lambda_sparse	Weight of sparsity regularization	1 × 10−4, 1 × 10−3, 1 × 10−2	1 × 10−2

. The results showed that TabNet consistently achieved high AUC values across different hyperparameter combinations, with an average AUC of 0.9759 and a standard deviation of only 0.0033. This indicates that TabNet’s performance remains stable and robust under a wide range of configurations, and that its parameter settings offer strong representativeness for structured geospatial data.

Although TabNet achieved excellent performance in this study, its generalizability remains to be further validated. Future research should focus on evaluating the model across larger, more complex datasets and diverse geographic regions to assess its robustness under varying environmental and geological conditions. In addition, due to its relatively high computational cost during training, future work could explore model optimization strategies to improve efficiency and enhance applicability to large-scale or real-time scenarios.

5.3. Relevance and Limitations of Susceptibility Indicators

Through feature importance analysis of the TabNet model, it was found that elevation, land cover, and annual NDVI play a dominant role in the model. This aligns with theoretical knowledge in the field of geology and supports findings from previous research [69]. However, some studies have reached different conclusions. For instance, some studies emphasize the significant role of rainfall in landslide susceptibility, suggesting that precipitation is a key factor in determining landslide occurrence [70]. These differences may be attributed to the adoption of different modeling approaches, which vary in how they interpret and assign importance to input features. In addition, regional differences in topography, geological conditions, and climate can further influence the relevance of specific factors. Therefore, future landslide susceptibility modeling requires the selection of an appropriate model and careful consideration of the study area’s environmental characteristics. Given the high importance of land cover revealed in this study, future research could further explore human–environment interactions by integrating indicators of anthropogenic activity. Examples include night-time light intensity, population density, or infrastructure distribution, which may help better capture the impact of human disturbance on slope stability and support the development of more comprehensive, coupled modeling frameworks.

Another important aspect not considered in this study is the role of extreme weather events. Due to the focus of this study on long-term landslide susceptibility rather than event-based hazard assessment, it does not account for extreme weather events, such as short-term intense rainfall or typhoons. However, these events are widely recognized as critical triggers of landslides. Future research could address this limitation by incorporating high-resolution meteorological data or historical records of extreme events to enhance the model’s temporal prediction capability.

In addition, this study does not explicitly account for the spatial autocorrelation between landslide occurrences and explanatory variables, which may influence model predictions and statistical inference. In spatial phenomena such as landslides, nearby locations often share similar environmental characteristics, leading to the spatial clustering of both the dependent and independent variables. Ignoring this spatial dependence may result in biased estimation of feature importance or overestimation of model performance due to the violation of the independence assumption. Future research could incorporate spatial statistical methods or geospatial machine learning frameworks such as spatial lag models to better capture the influence of spatial dependence and improve the robustness of landslide susceptibility assessments.

6. Conclusions

This study constructs a series of landslide susceptibility prediction models based on slope unit partitioning and historical landslide data in Ya’an City. Comparative experiments are conducted from two key perspectives: negative sample construction strategies and modeling architectures, leading to the following main conclusions:

(1)

The GWR-based negative sampling method consistently outperforms traditional random sampling across all evaluated models, significantly improving both discriminative accuracy and generalization performance. By selecting negative samples from regions with low composite explanatory power, the GWR approach enhances the representativeness of low-risk areas and improves the distinction between positive and negative instances.

(2)

As a deep neural network specifically designed for structured data, TabNet achieves outstanding performance in landslide susceptibility modeling, attaining an average AUC of 0.9828, which is substantially higher than those of RF, LightGBM, DNN, and ResNet. TabNet exhibits strong adaptability to structured inputs, high predictive accuracy, and robust interpretability, confirming its effectiveness and suitability for modeling complex, multi-factor geological hazards.

(3)

Feature importance analysis identifies elevation, land cover, and annual NDVI as the most highly influential factors, highlighting the combined effects of topographic conditions, anthropogenic disturbances, and vegetation cover on landslide susceptibility.