Comparative assessment of landslide susceptibility in Fugu town using machine learning models at multiple grid resolutions Authors: Shuhang Ni, Yang Xie Published: 2025 ABSTRACT Landslides pose serious threats to life and infrastructure in the loess hilly regions of northern China, where the complex interaction between topography, lithology, and human activities makes susceptibility prediction challenging. Although numerous studies have explored the effects of spatial resolution and machine learning models on landslide susceptibility assessment, their conclusions remain inconsistent, and few have quantitatively examined the interaction between these two factors. To address this gap, this study conducted a multi-scale comparative analysis of landslide susceptibility in Fugu Town, Shaanxi Province (area: 328.73 km2) using Random Forest (RF) and XGBoost (XGB) models under 15 m, 30 m, and 60 m grid resolutions. The objectives were to determine the optimal model–resolution combination and to evaluate the influence of spatial scale on model performance. Results indicate that grid resolution significantly affects prediction accuracy and spatial patterns of susceptibility zones. The XGBoost model at 15 m resolution achieved the best performance (AUC = 0.980), with high- and very-high-susceptibility zones mainly distributed in the canyon-hill terrains of the northern and central-western regions. These findings provide methodological insights for selecting appropriate model–resolution combinations and practical references for landslide risk management in the loess areas of northern Shaanxi. Introduction Landslides are among the most frequent and destructive geological hazards worldwide, posing severe threats to human life, infrastructure, and the environment. Their spatial distribution and developmental characteristics are influenced by both natural factors (such as topography, lithology, and hydrology) and human activities (such as excavation and construction). In recent years, rapid urbanization and infrastructure expansion in mountainous regions of China have intensified the frequency of slope failures, resulting in substantial casualties and economic losses. Therefore, reliable landslide susceptibility assessment (LSA) is essential for effective disaster prevention, regional planning, and land-use management. One of the key challenges in improving landslide susceptibility mapping lies in determining the appropriate spatial resolution for data representation. Grid resolution directly determines the scale and precision of terrain characterization, influencing information extraction, model input quality, and overall prediction accuracy. Variations in resolution can lead to substantial differences in the depiction of terrain morphology, relief intensity, drainage networks, and slope structures, thus altering the outcomes of susceptibility evaluation. Consequently, identifying the impact of spatial resolution on modeling performance has become a central focus in geohazard prediction research. Previous research has extensively examined the influence of spatial resolution on landslide susceptibility modeling, but the reported findings are often inconsistent. For instance, Huang et al.1observed that model accuracy decreases with coarser resolutions, achieving optimal results at 15 m, while Cavazzi et al.2 argued that finer resolution does not always ensure better predictions. Cama et al.3 and Shirzadi et al.4 found that the performance of susceptibility models varies significantly with grid scale, whereas Schlogel et al.5 and Palamakumbure et al.6 reported optimal resolutions between 10–25 m and 10–40 m, respectively. Conversely, Azemeraw 7obtained superior results at 90 m resolution for large-scale landslides, and Tian et al.8 suggested that excessively high resolutions may even degrade mapping accuracy in urban settings. These conflicting conclusions indicate that the optimal grid resolution depends strongly on both the landslide scale and the geomorphological setting, thereby emphasizing the need for multi-scale comparative analyses in different regional contexts. Although several studies have reported that model accuracy tends to decline with coarser resolution, others have found contrasting results depending on geomorphological settings, data characteristics, and model types. Therefore, it remains necessary to perform a systematic multi-scale analysis to better understand these relationships under the unique loess hilly conditions of northern Shaanxi. However, the accuracy of susceptibility mapping is not solely determined by spatial resolution; the selection of predictive models also plays a decisive role. In recent years, data-driven machine learning9,10 algorithms—such as logistic regression11, artificial neural networks12, support vector machines13, fuzzy logic14, decision trees15, Random Forest (RF)16, and XGBoost (XGB)17—have been widely adopted due to their strong adaptability and ability to handle nonlinear relationships. Despite these advances, most studies have focused on either the effect of resolution or model performance independently. Few have systematically explored how the interaction between spatial resolution and model type influences prediction accuracy. This lack of integrated analysis limits our understanding of how data scale and model structure jointly determine the reliability of susceptibility assessment results. Among various machine learning algorithms, Random Forest (RF) and XGBoost (XGB) were selected in this study due to their proven robustness, interpretability, and superior performance in nonlinear classification problems. Traditional approaches such as logistic regression (LR) and support vector machines (SVM) often suffer from limited capacity in capturing complex feature interactions or require extensive preprocessing of non-linear data. Artificial neural networks (ANN), although powerful, are prone to overfitting and demand large sample sizes and intensive computational resources. RF, as an ensemble tree-based method, performs well with noisy and high-dimensional datasets, providing variable importance measures that facilitate model interpretation. XGBoost, an optimized gradient boosting algorithm, improves upon RF by sequentially correcting residual errors and incorporating regularization to prevent overfitting. Its efficiency and accuracy have been widely demonstrated in recent geohazard prediction studies. Therefore, RF and XGBoost were chosen to represent two representative ensemble learning paradigms for comparative landslide susceptibility analysis. To address these research gaps, this study performs a comparative landslide susceptibility assessment in Fugu Town, northern Shaanxi Province, using Random Forest (RF) and XGBoost (XGB) models under three grid resolutions (15 m, 30 m, and 60 m). The main objectives are to: (1) quantitatively evaluate the influence of grid resolution on susceptibility mapping accuracy; (2) compare the predictive performance of RF and XGBoost models at multiple spatial scales; and (3) identify the optimal model–resolution combination suitable for loess hilly terrains. The findings of this study provide methodological insights into the scale–model relationship in landslide susceptibility assessment and offer practical guidance for regional disaster prevention and geohazard risk management in the loess areas of northern Shaanxi. Study area and dataset Study area Fugu Town is situated in the southeastern part of Fugu County, Yulin City, Shaanxi Province, with geographical coordinates ranging from 110°53′38″E to 110°06′32″E and 38°59′42″N to 39°06′30″N (Fig. 1). The town spans an area of 328.73 km2.The study area lies at the transitional boundary between the Inner Mongolia Plateau and the northeastern Loess Plateau of Shaanxi, structurally positioned on a westward-dipping monocline. From west to east, it traverses three geomorphological units: loess ridge-gully areas, canyon-hill areas, and fluvial terraces. The loess ridge-gully areas are characterized by homogeneous lithology and high permeability, with limited catchment capacity. During heavy rainfall events, surface runoff rapidly erodes soil, leading to severe water loss and soil erosion. Canyon-hill landscapes, primarily distributed along river valleys, exhibit steep slopes and exposed bedrock due to intense loess denudation. These areas are prone to rockfalls and landslides owing to unstable slopes and precipitous terrain. Fluvial terraces, elevated 5–10 m above the Yellow River channel, consist of Quaternary alluvial deposits.Fig. 1Geographical location and landslide distribution. Geographical location and landslide distribution. The stratigraphy includes Ordovician, Carboniferous, Permian, Triassic, Jurassic, Neogene, and Quaternary formations. Mudstone and clay-rich strata dominate, exhibiting high hydrophilicity. When saturated with groundwater, these layers form weak structural planes with reduced shear strength, significantly elevating landslide susceptibility. Major transportation infrastructure—such as the Shenfu Expressway, Fu-Zhun Highway, Yanhuang Highway, and Shenshuo Railway—crosses the region. Slope excavation activities during road and railway construction have created high-steep slopes. Coupled with the interbedded sandstone-mudstone lithology, which exhibits contrasting mechanical properties, slope instability frequently triggers collapse and landslide events. Data preparation Data sources The landslide susceptibility assessment was conducted using multisource geospatial data, including airborne-derived DEMs at 15 m, 30 m, and 60 m resolutions, ASTER GDEM, and a 1:10,000-scale geological map of Fugu Town. Landslide Inventory A systematic landslide inventory, comprising 254 recorded events (225 collapses, 27 landslides, and 2 debris flows), was established through large-scale field surveys (1:5,000) and integration of geological archives. Collapses dominated the inventory (88.6%), primarily small-scale rockfalls (< 5,000 m3), while landslides were characterized as shallow soil slips (< 3 m depth) on 25–45° slopes, and debris flows occurred in first-order drainage basins (< 2 km2). Factor selection and analysis Eight susceptibility factors—slope, slope aspect, curvature, degree of relief, lithology, slope structure, distance from roads, and distance from rivers—were selected based on hydrogeological mechanisms (Table 1). Multi-scale factor extraction was achieved through 15 m DEM resampling and spatial analysis (Fig. 2), ensuring dimensional consistency across evaluation layers.Table 1Classification criteria for evaluation factors.FactorsCategoriesSlope(°)(1) 0–5; (2) 5–10; (3) 10–15;(4) 15–20;(5) > 25Slope aspectFlat;(2) North;(3) Northeast;(4) East; (5) Southeast; (6) South;(7) Southwest; (8) West; (9) NorthwestCurvature(1) < -0.2;(2)-0.2–0.1;(3)-0.1–0;(4)0–0.1;(5)0.1–0.2;(6) > 0.2Terrain relief(1) < 60;(2)60–80;(3)80–100;(4)100–120;(5) > 120Distance to roads(m)(1) < 40;(2)40–80;(3)80–120;(4)120–160;(5) > 160Distance to rivers(m)(1) < 100;(2)100–200;(3)200–300;(4)300–400;(5) > 400Lithology(1) Hard-semi-hard clastic rock assemblage;(2) Soft sandy clay clastic rock assemblage;(3) Coarse sand gravel and loess soil double soil;(4) loess;(5) Hard carbonate rock assemblageSlope structure(1) Dip slope;(2) Downslope;(3) Cross slope;(4) Upslope;(5) Strike slopeFig. 2Spatial distribution of evaluation factor layers. (a) slope, (b) slope aspect, (c) curvature, (d) degree of relief, (e) lithology, (f) slope structure, (g) distance from roads, (h) distance from rivers. Table 1Classification criteria for evaluation factors.FactorsCategoriesSlope(°)(1) 0–5; (2) 5–10; (3) 10–15;(4) 15–20;(5) > 25Slope aspectFlat;(2) North;(3) Northeast;(4) East; (5) Southeast; (6) South;(7) Southwest; (8) West; (9) NorthwestCurvature(1) < -0.2;(2)-0.2–0.1;(3)-0.1–0;(4)0–0.1;(5)0.1–0.2;(6) > 0.2Terrain relief(1) < 60;(2)60–80;(3)80–100;(4)100–120;(5) > 120Distance to roads(m)(1) < 40;(2)40–80;(3)80–120;(4)120–160;(5) > 160Distance to rivers(m)(1) < 100;(2)100–200;(3)200–300;(4)300–400;(5) > 400Lithology(1) Hard-semi-hard clastic rock assemblage;(2) Soft sandy clay clastic rock assemblage;(3) Coarse sand gravel and loess soil double soil;(4) loess;(5) Hard carbonate rock assemblageSlope structure(1) Dip slope;(2) Downslope;(3) Cross slope;(4) Upslope;(5) Strike slope Classification criteria for evaluation factors. Flat;(2) North;(3) Northeast;(4) East; (5) Southeast; (6) South; (7) Southwest; (8) West; (9) Northwest Spatial distribution of evaluation factor layers. (a) slope, (b) slope aspect, (c) curvature, (d) degree of relief, (e) lithology, (f) slope structure, (g) distance from roads, (h) distance from rivers. Slope Slope, defined as the steepness of terrain units, is a critical parameter in slope stability analysis. It directly influences gravitational differentiation and shear strength of slope materials. When slope gradients exceed critical thresholds, gravitational shear stress surpasses the material’s shear strength, triggering instability. In various mathematical models, slope gradient consistently demonstrates high weighting coefficients and contribution rates, with applicability unrestricted by regional geological settings. Therefore, slope gradient was selected as a key evaluation factor for the study area. Slope aspect Slope aspect governs solar radiation exposure, thereby regulating weathering and erosion processes. Weathering rates vary significantly across different slope aspects: sun-facing slopes exhibit intense physical weathering due to diurnal temperature fluctuations, resulting in highly fragmented surface materials, whereas shade-facing slopes are dominated by chemical weathering, characterized by elevated clay mineral content, reduced permeability, and the formation of perched water zones during rainfall. These dynamics exert substantial impacts on slope stability. Curvature Curvature quantifies the concavity or convexity of surface geometry, directly modulating internal stress distribution and deformation patterns within slopes. Planar curvature reflects lateral bending, while profile curvature describes longitudinal undulations. High-curvature zones concentrate stresses, accelerating fracture propagation in surface materials. Additionally, curvature influences surface runoff pathways and infiltration rates, thereby affecting slope stability. Concave slopes (negative curvature) promote water accumulation, increasing pore pressure and weakening shear strength, whereas convex slopes (positive curvature) intensify erosional runoff, destabilizing slope toes. The study area’s prevalence of high convex curvature critically enhances landslide susceptibility. Degree of relief Degree of relief measures the relative elevation difference within a specific area, reflecting variations in surface elevation and terrain complexity. Empirical studies confirm a significant positive correlation between relief intensity and disaster frequency. Furthermore, relief interacts synergistically with slope gradient to establish multi-scale hazard prediction frameworks. The study area contains overlapping zones of steep slopes and high relief, amplifying landslide risks. Slope structure Slope structure significantly controls the spatial distribution and failure modes of landslides. Determined by the relationship between slope aspect and bedrock dip direction, it reflects stress environments at hazard locations. In Fugu Town’s NW-trending monoclinal structure (dip direction: 270–360°), field-measured bedding orientation data were interpolated using Kriging. Absolute differences between slope aspect and dip direction were calculated and classified into five structural types: cataclinal, anaclinal, orthoclinal, subcataclinal, and subanaclinal slopes. Lithology Lithology dictates material sources for landslides and controls their types and scales. Four lithological units dominate Fugu Town: (1) layered hard-to-semi-hard clastic rocks, (2) layered weak sandy clay clastics, (3) coarse sand-gravel loess double-layer soils, and (4) layered hard carbonate rocks. The predominant clastic group weathers into red clay and mudstone—materials with loose structures, poor water retention, and low erosion resistance, predisposing slopes to collapse under rainfall. Distance from rivers Fluvial erosion during riverflow deepens and widens channels, destabilizing adjacent slopes. Consequently, riverbanks are landslide-prone zones. The study area’s dense river network, including the Yellow River, exacerbates these risks, particularly during monsoons. Proximity to rivers was thus identified as a critical susceptibility factor. Distance from roads Slope excavation during road construction generates high-steep cuts, disrupting inherent stress equilibria and elevating landslide probabilities. Mountain highway development directly influences landslide spatial patterns. Therefore, proximity to roads was included as a key conditional factor in susceptibility assessment. Methodology This study was designed to generate enhanced landslide susceptibility zonation for Fugu County through comparative analysis of two machine learning algorithms—Random Forest (RF) and XGBoost—under multi-scale grid resolutions (15 m, 30 m, 60 m). (Fig. 3) presents the methodological flowchart, outlining key procedures undertaken in the study: (1) landslide inventory compilation, (2) preparation of conditioning factors, (3) model evaluation and susceptibility zonation mapping, (4) comparative performance assessment, and (5) optimal model selection for final susceptibility classification.Fig. 3Flow chart of the study. Flow chart of the study. Geological hazard susceptibility modeling and evaluation methods Modeling methods Random forest model The Random Forest (RF) algorithm18, an advanced machine learning method based on ensemble learning, was employed to enhance prediction accuracy and generalization capabilities through the construction and integration of multiple decision trees. The bagging technique was first implemented, where random sampling with replacement (typically selecting two-thirds of the original training data) was conducted to generate numerous independent training subsets. For each subset, a separate decision tree was developed. During tree construction, additional stochasticity was introduced by randomly selecting feature subsets at each node for splitting, a strategy that effectively reduced overfitting risks while improving model robustness19,20. Following the development of multiple decision trees, predictions were aggregated through ensemble learning: majority voting was adopted for classification tasks, whereas regression outputs were determined by averaging predictions across all trees. Notably, the remaining one-third unsampled data (out-of-bag, OOB) were strategically utilized for internal error estimation. The OOB error, calculated as the mean prediction error across all trees, provided a reliable estimate of generalization error. This mechanism not only facilitated effective model performance evaluation but also enhanced the algorithm’s practical utility and reliability. The mathematical expression is presented as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c}{P}^{*}\le \frac{\rho \left(1-{S}^{2}\right)}{{S}^{2}}\#\end{array}$$\end{document} In the formulation P* represents the generalization error of the Random Forest (RF) model ρ denotes the average correlation coefficient among decision trees s indicates the average strength of individual decision trees. As demonstrated in Equation, the generalization capability of the RF model can be enhanced through two optimization pathways: (1) reducing the average inter-tree correlation coefficient (ρ), or (2) increasing the average tree strength (s). This dual mechanism fundamentally explains RF’s superior resistance to overfitting compared to single decision tree classifiers in landslide susceptibility modeling. XGBoost model The XGBoost algorithm, an advanced gradient boosting technique21, represents an optimized iteration of the Gradient Boosting Decision Tree (GBDT) framework. Built upon an additive forward modeling architecture, this method strategically integrates multiple weak learners (decision trees) to construct a robust ensemble predictor. In operational terms, input samples were sequentially processed through successive decision trees, with predictions accumulated additively to generate final outputs. A critical innovation lies in its residual-driven tree construction: each subsequent tree was designed to explicitly address prediction residuals from preceding models, thereby systematically reducing model bias. To mitigate overfitting risks, a regularization term was incorporated into the objective function to penalize excessive tree complexity. This dual optimization mechanism—residual correction and regularization—significantly enhanced the model’s generalization capacity. The XGBoost objective function ​is mathematically defined as: Objective Function = Loss Function + Regularization Term.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c}ob{j}^{\left(t\right)}={\sum }_{i=1}^{m}L\left({y}_{i},{\widehat{y}}_{i}^{\left(t-1\right)}+{f}_{t}\left({x}_{i}\right)\right)+{\sum }_{r=1}^{t}\Omega \left({f}_{r}\right)\#\end{array}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c}\Omega \left({f}_{t}\right)=\gamma T+\frac{1}{2}\lambda {\sum }_{j=1}^{T}{\omega }_{j}^{2}\#\end{array}$$\end{document} In the formulation:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L\left({y}_{i},{\widehat{y}}_{i}^{\left(t-1\right)}+{f}_{t}\left({x}_{i}\right)\right)$$\end{document} is defined as the loss function for training samples, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{i}$$\end{document} represents the observed value of the i-th sample, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{i}^{{\left( {t - 1} \right)}}$$\end{document} denotes the predicted value after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {t - 1} \right)$$\end{document} iterations.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{t}\left({x}_{i}\right)$$\end{document} corresponds to the decision tree function generated during the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}-th training cycle for sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i}$$\end{document}.The regularization term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Omega \left({f}_{t}\right)=\gamma T+\frac{1}{2}\lambda {\sum }_{j=1}^{T}{\omega }_{j}^{2}$$\end{document} quantifies model complexity through:o\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}: Number of nodes in the t-th treeo\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{j}^{{}}$$\end{document}: Weight value at the j-th nodeo\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda$$\end{document}: Hyperparameters controlling regularization intensity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L\left({y}_{i},{\widehat{y}}_{i}^{\left(t-1\right)}+{f}_{t}\left({x}_{i}\right)\right)$$\end{document} is defined as the loss function for training samples, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{i}$$\end{document} represents the observed value of the i-th sample, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{i}^{{\left( {t - 1} \right)}}$$\end{document} denotes the predicted value after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {t - 1} \right)$$\end{document} iterations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L\left({y}_{i},{\widehat{y}}_{i}^{\left(t-1\right)}+{f}_{t}\left({x}_{i}\right)\right)$$\end{document} is defined as the loss function for training samples, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{i}$$\end{document} represents the observed value of the i-th sample, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_{i}^{{\left( {t - 1} \right)}}$$\end{document} denotes the predicted value after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left( {t - 1} \right)$$\end{document} iterations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{t}\left({x}_{i}\right)$$\end{document} corresponds to the decision tree function generated during the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}-th training cycle for sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{t}\left({x}_{i}\right)$$\end{document} corresponds to the decision tree function generated during the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t$$\end{document}-th training cycle for sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i}$$\end{document}. The regularization term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Omega \left({f}_{t}\right)=\gamma T+\frac{1}{2}\lambda {\sum }_{j=1}^{T}{\omega }_{j}^{2}$$\end{document} quantifies model complexity through:o\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}: Number of nodes in the t-th treeo\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{j}^{{}}$$\end{document}: Weight value at the j-th nodeo\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda$$\end{document}: Hyperparameters controlling regularization intensity The regularization term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Omega \left({f}_{t}\right)=\gamma T+\frac{1}{2}\lambda {\sum }_{j=1}^{T}{\omega }_{j}^{2}$$\end{document} quantifies model complexity through:o\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}: Number of nodes in the t-th treeo\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{j}^{{}}$$\end{document}: Weight value at the j-th nodeo\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda$$\end{document}: Hyperparameters controlling regularization intensity o\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}: Number of nodes in the t-th tree \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T$$\end{document}: Number of nodes in the t-th tree o\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{j}^{{}}$$\end{document}: Weight value at the j-th node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega_{j}^{{}}$$\end{document}: Weight value at the j-th node o\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda$$\end{document}: Hyperparameters controlling regularization intensity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda$$\end{document}: Hyperparameters controlling regularization intensity This composite expression is formally designated as the ​scoring function, serving as an evaluative framework to quantify the structural quality of decision trees. Specifically, ​a lower score indicates superior tree architecture in meeting optimization objectives, where minimization simultaneously achieves: (1) Enhanced prediction accuracy (via loss reduction), (2) Optimal model simplicity (through regularization constraints). Model optimization and hyperparameter tuning To ensure fair and optimized model comparison, hyperparameter tuning was conducted for both RF and XGBoost models. for RF, a grid search approach was adopted to optimize the number of trees, maximum tree depth, and number of features considered per split, with performance evaluated using out-of-bag (OOB) error estimation. for XGBoost, parameters such as learning rate, maximum depth, subsample ratio, and number of boosting rounds were fine-tuned using tenfold cross-validation combined with early stopping to prevent overfitting. The parameter ranges were determined empirically based on preliminary trials, and final configurations were selected according to the highest mean AUC value on the validation dataset. Validation methods for susceptibility evaluation results The validation of susceptibility evaluation results ​was recognized as a critical process for assessing model performance and prediction accuracy. This systematic verification ​served dual purposes: interpreting the model’s predictive capacity and providing constructive feedback for iterative model refinement. The validation ​was implemented at two distinct analytical levels: (1) Rationality verification of susceptibility zoning: The spatial coherence of susceptibility zoning outcomes ​was examined by comparing the classified susceptibility levels (e.g., high, moderate, low) with the actual distribution of historical landslides within the study area. (2)Quantitative accuracy assessment: The predictive precision ​was rigorously evaluated using the Area Under the Curve (AUC) metric22 derived from Receiver Operating Characteristic (ROC) analysis. Rationality verification The rationality verification of landslide susceptibility zoning was conducted by establishing a multi-dimensional statistical indicator system to validate the spatial consistency between model predictions and observed hazard distributions. The validation process followed the synergistic relationship between zoning area and hazard point density, employing quantitative metrics to reveal spatial differentiation characteristics of risk levels. Key metrics included: Zoning area proportion: Representing the spatial scale of susceptibility levels (e.g., high/medium/low); Hazard point proportion: Reflecting hazard concentration across zones; Hazard point density (number of hazards per unit area): Quantifying disaster intensity per spatial unit. The core logic focused on verifying whether high-susceptibility zones exhibited significant hazard clustering effects, adhering to the statistical principle of ​low spatial coverage but high hazard density. For instance, when the area proportion of very-high/high susceptibility zones was < 30% while their hazard density exceeded three times that of medium–low susceptibility zones, the zoning results were deemed consistent with the spatial selectivity of landslide mechanisms. Tabular data were further analyzed using parameters such as ​frequency ratio and ​information value to quantitatively characterize the significance of hazard development across zones. ROC curve The ROC (Receiver Operating Characteristic) curve was utilized as a statistical tool to quantify model classification performance in landslide susceptibility assessment. By plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) across classification thresholds, the discriminative capability of the model for distinguishing hazard-prone from stable areas was systematically assessed. The ​AUC (Area Under the Curve) value, ranging from 0.5 (random classification) to 1.0 (perfect prediction), served as the primary accuracy metric. Higher AUC values indicated stronger model interpretability of spatial hazard differentiation. A key advantage of ROC analysis lies in its ​robustness to imbalanced sample distributions, ensuring reliable evaluation even when sparse hazard units (positive samples) coexist with extensive non-hazard units (negative samples). Dynamic threshold adjustment balanced sensitivity (TPR) and specificity (1 − FPR), thereby mitigating model bias toward majority classes. Furthermore, the ROC curve enabled ​cross-model comparisons and ​threshold optimization. The Youden Index (J = TPR − FPR) was applied to identify optimal thresholds that maximized classification efficacy, facilitating precise delineation of high-, medium-, and low-susceptibility zones. This probabilistic framework provided actionable spatial guidance for landslide risk management. Feature importance evaluation To quantitatively evaluate the relative contribution of each conditioning factor to landslide susceptibility, the Information Value (IV) method was employed. This method measures the statistical dependence between the presence of landslides and the classification intervals of each factor by computing the information content of each class. The IV value reflects the strength and direction of the relationship between a conditioning factor and landslide occurrence. A higher IV value indicates a stronger correlation and greater influence on landslide formation. For each grid resolution (15 m, 30 m, and 60 m), the information value was calculated for every class of all conditioning factors. The maximum IV value of each factor was used as an indicator of its relative importance in controlling landslide susceptibility. This approach provides an independent and interpretable way to identify key variables without relying solely on the internal feature-importance metrics of machine learning models. Prediction of landslide susceptibility Correlation analysis of evaluation factors The Pearson correlation coefficient23,24 was applied to quantify linear dependencies among evaluation factors, calculated as the covariance of variables X and Y divided by the product of their standard deviations. For paired samples (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{i}$$\end{document},\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Y}_{i}$$\end{document}) with means \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\text{X} }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{\text{Y} }$$\end{document}, the coefficient r is expressed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r = \frac{{\sum\limits_{i = 1}^{n} {\left( {X_{i} - \overline{X}} \right)\left( {Y_{i} - \overline{Y}} \right)} }}{{\sqrt {\sum\limits_{i = 1}^{n} {\left( {X_{i} - \overline{X}} \right)^{2} } } \sqrt {\sum\limits_{i = 1}^{n} {\left( {Y_{i} - \overline{Y}} \right)^{2} } } }}$$\end{document}where n is the sample size. Values of r range from -1 (perfect negative correlation) to + 1 (perfect positive correlation), with magnitudes interpreted as: ∣r∣ < 0.3 (negligible), 0.3 ≤ ∣r∣ ≤ 0.5 (low), 0.5 < ∣r∣ ≤ 0.8 (moderate), and ∣r∣ > 0.8 (high). As shown in Fig. 4, all eight factors exhibited weak inter-correlations (∣r∣ < 0.5), confirming their statistical independence and suitability for susceptibility modeling without multicollinearity risks.Fig. 4Factor correlation heatmap. Factor correlation heatmap. Landslide susceptibility mapping This study utilized Random Forest (RF) and XGBoost models to predict landslide susceptibility in Fugu Town under three grid resolutions (15 m, 30 m, and 60 m). The results were classified into five susceptibility levels (very low, low, moderate, high, and very high) using the natural breaks method, generating a total of six landslide susceptibility maps (Fig. 5).Fig. 5Susceptibility zoning map (a) 15mRF (b) 15mXGB (c) 30mRF (d) 30mXGB (e) 60mRF (f) 60mXGB. Susceptibility zoning map (a) 15mRF (b) 15mXGB (c) 30mRF (d) 30mXGB (e) 60mRF (f) 60mXGB. The susceptibility zoning maps indicate that the spatial distribution of susceptibility levels is generally consistent across different models and resolutions. High and very high susceptibility zones are primarily located in the canyon-hill terrain of northern and central-western Fugu Town, while large areas of low and very low susceptibility zones are distributed in the loess ridge-gully terrain in the western region. These high-to-extreme susceptibility zones are strongly associated with slopes > 20°, terrain relief of 60–80 m, proximity to roads (40–80 m buffer), and hard to semi-hard clastic rock formations. From the perspective of spatial resolution, susceptibility maps generated at the same resolution exhibit high similarity. However, significant differences are observed when comparing maps across resolutions: the 15 m and 30 m resolution maps show distinct spatial pattern variations compared to the 60 m resolution maps. Regarding model performance, comparisons of susceptibility zoning maps under identical resolutions reveal that the Random Forest model assigns higher susceptibility levels and produces broader coverage of high-to-extreme susceptibility zones than the XGBoost model. Furthermore, the RF model consistently predicts elevated susceptibility levels in the same geographic areas compared to XGBoost. Results validation Rationality verification The susceptibility zoning results across different grid resolutions (Table 2) demonstrate that as the susceptibility level increases, the corresponding area proportion gradually decreases, while the distribution density of hazard sites within these zones significantly rises. This inverse relationship between susceptibility level magnitude and spatial coverage, coupled with the concentration of hazards in high-risk zones, validates the rationality and reliability of the susceptibility classification framework proposed in this study.Table 2Susceptibility zoning results.ResolutionmodelSusceptibilitylevelSusceptibilityArea(km2)Area percentage(%)Disaster quantityDisasterpercentage(%)DisasterDensity(/10km2)15mRFVery low101.1130.7631.180.30Low79.1924.09114.331.39Moderate55.7616.964417.327.89High55.1016.766124.0211.07Very high37.5711.4313553.1535.93XGBVery low124.5938.0720.790.16Low62.3219.04155.912.41Moderate49.5015.12249.454.85High49.1615.024216.548.54Very high41.7212.7517167.3240.9830mRFVery low114.3534.79124.721.05Low68.1420.73197.482.79Moderate54.9116.714015.757.28High55.8617.005220.479.31Very high35.4410.7813151.5736.97XGBVery low142.0443.08145.510.99Low69.2021.05239.063.32Moderate51.6515.713413.396.58High36.2511.034116.1411.31Very high30.599.3014255.9146.4260mRFVery low83.9625.53207.872.38Low76.5923.293011.813.92Moderate69.8021.235120.087.31High53.7716.356927.1712.83Very high44.7013.598433.0718.79XGBVery low162.3749.4020.790.12Low47.1614.35103.942.12Moderate33.8410.29239.066.80High35.4310.784517.7212.70Very high49.9015.1817468.5034.87 Table 2Susceptibility zoning results.ResolutionmodelSusceptibilitylevelSusceptibilityArea(km2)Area percentage(%)Disaster quantityDisasterpercentage(%)DisasterDensity(/10km2)15mRFVery low101.1130.7631.180.30Low79.1924.09114.331.39Moderate55.7616.964417.327.89High55.1016.766124.0211.07Very high37.5711.4313553.1535.93XGBVery low124.5938.0720.790.16Low62.3219.04155.912.41Moderate49.5015.12249.454.85High49.1615.024216.548.54Very high41.7212.7517167.3240.9830mRFVery low114.3534.79124.721.05Low68.1420.73197.482.79Moderate54.9116.714015.757.28High55.8617.005220.479.31Very high35.4410.7813151.5736.97XGBVery low142.0443.08145.510.99Low69.2021.05239.063.32Moderate51.6515.713413.396.58High36.2511.034116.1411.31Very high30.599.3014255.9146.4260mRFVery low83.9625.53207.872.38Low76.5923.293011.813.92Moderate69.8021.235120.087.31High53.7716.356927.1712.83Very high44.7013.598433.0718.79XGBVery low162.3749.4020.790.12Low47.1614.35103.942.12Moderate33.8410.29239.066.80High35.4310.784517.7212.70Very high49.9015.1817468.5034.87 Susceptibility zoning results. Accuracy validation After calculating the landslide susceptibility indices under six different model-resolution combinations, the Receiver Operating Characteristic (ROC) curve25,26 (Fig. 6) and its Area Under the Curve (AUC) values were employed to evaluate prediction accuracy. Comprehensive analysis of the validation results reveals the following:Fig. 6ROC curve (a) RF model (b) XGBoost model. ROC curve (a) RF model (b) XGBoost model. Random Forest (RF) model: AUC values for 15 m, 30 m, and 60 m resolutions are 0.908, 0.894, and 0.864, respectively. XGBoost model: AUC values for corresponding resolutions are 0.967, 0.956, and 0.945. All AUC values exceed 0.85, indicating high overall prediction accuracy. Notably, AUC values decrease as grid resolution coarsens (from 15 to 60 m), with the 15 m resolution achieving the highest accuracy, thus identified as the optimal resolution in this study. Furthermore, under identical resolutions, XGBoost consistently outperforms RF in AUC metrics, demonstrating its superior predictive capability for landslide susceptibility in Fugu Town. Feature importance based on information value The information value analysis revealed distinct variations in factor importance under different grid resolutions (Table 3). Among all conditioning factors, slope, distance to roads, curvature, slope structure type, and aspect exhibited the highest maximum information values, indicating that these variables are the key factors inducing geological disasters in the study area. These parameters are directly related to slope gradient, gravitational stress, and human disturbance intensity, which jointly control slope stability.Table 3IV of conditioning factors under different grid resolutions.Conditioning FactorIV (15 m)IV (30 m)IV (60 m)Slope (°)0.6850.6420.598Distance to roads (m)0.5120.4890.466Curvature0.4380.4150.372Slope structure type0.4020.3890.355Aspect0.3660.3520.327Terrain relief (m)0.3180.2960.271Distance to rivers (m)0.2870.2750.251Lithology0.2450.2310.218 Table 3IV of conditioning factors under different grid resolutions.Conditioning FactorIV (15 m)IV (30 m)IV (60 m)Slope (°)0.6850.6420.598Distance to roads (m)0.5120.4890.466Curvature0.4380.4150.372Slope structure type0.4020.3890.355Aspect0.3660.3520.327Terrain relief (m)0.3180.2960.271Distance to rivers (m)0.2870.2750.251Lithology0.2450.2310.218 IV of conditioning factors under different grid resolutions. In contrast, terrain relief, distance to rivers, and lithology showed relatively smaller IV values, implying that their influence on landslide occurrence is secondary. The variation of IV values with resolution also indicates that finer grids (15 m) tend to strengthen the explanatory power of topographic factors, while coarser grids (60 m) emphasize broader geomorphic controls. Discussion Impact of spatial resolution on landslide susceptibility assessment From a spatial resolution perspective, the 15 m grid exhibited the best overall performance in this study, which is consistent with several previous findings Lee27, Garosi et al.28. However, finer resolution does not always guarantee higher prediction accuracy. Excessively detailed DEMs can amplify local noise, increase data redundancy, and potentially distort slope or curvature parameters in flat terrains Claessens et al.29. Moreover, Hong H30 emphasized that the optimal grid size depends on both the geomorphological context and the scale of slope failures. In this study, while small-scale landslides were better captured at 15 m, a large collapse along the Yellow River Highway was more accurately represented at 30 m, confirming that the best-performing resolution varies with landslide type and terrain heterogeneity. Therefore, the determination of an “optimal” resolution should balance spatial detail, noise sensitivity, and computational efficiency through multi-scale testing. Impact of evaluation models on landslide susceptibility assessment Both Random Forest (RF) and XGBoost models achieved AUC values exceeding 0.85 across all resolutions, with AUC > 0.8 generally indicating robust performance. While the two models operate at comparable tiers, XGBoost outperformed RF in Fugu Town’s susceptibility assessment. Data from the results table highlight XGBoost’s superiority: under all three resolutions, its proportions of hazard points in very high susceptibility zones (67.32%, 55.91%, and 68.50%) significantly surpassed RF’s values (53.15%, 51.57%, and 33.07%). This evidence confirms XGBoost’s greater suitability for landslide susceptibility modeling in Fugu Town, particularly in identifying high-risk areas. Conclusions This study conducted landslide susceptibility assessments in Fugu Town using Random Forest (RF) and XGBoost models at 15 m, 30 m, and 60 m spatial resolutions, followed by rationality verification and accuracy validation. The key conclusions are as follows: The spatial resolution of grid units significantly influences susceptibility assessment outcomes. AUC values increase with higher spatial resolutions, peaking at 15 m resolution for both RF (0.908) and XGBoost (0.980). At 15 m, 30 m, and 60 m resolutions, the AUC values for RF are 0.908, 0.894, and 0.875, respectively, while those for XGBoost reach 0.980, 0.956, and 0.954. These results confirm XGBoost’s superior predictive accuracy over RF under identical resolution conditions. High to very high susceptibility zones are predominantly concentrated in northern and central-western Fugu Town, characterized by slopes > 20°, terrain relief of 60–80 m, proximity to roads (40–80 m buffer), and lithology dominated by hard to semi-hard clastic rock formations.