Optimized Boosting Algorithm Accurately Maps Landslide Hazards in Maiji District

A geological modeling investigation conducted in the Maiji District of Tianshui City, China, demonstrated that integrating stratified non-landslide sampling based on engineering geological petrofabric with a particle swarm optimization-tuned Categorical Boosting algorithm produces an accurate spatial evaluation of regional landslide susceptibility. Landslides represent widespread, destructive geological disasters driven by endogenous geological structures and external environmental conditions, threatening infrastructure engineering and human settlements across the region. Regional landslide susceptibility mapping establishes spatial databases and multi-type indicator systems, combining statistical methods and machine learning architectures to predict the spatial probability of slope failures. While traditional heuristic assessments and early machine learning techniques such as Naive Bayes, Decision Trees, and Support Vector Machines have been employed historically, expanding spatial database volumes and multi-dimensional factors have increasingly challenged their predictive stability. Consequently, ensemble boosting frameworks that construct strong classifiers from iterative weak learners have become prominent tools for hazard prevention, disaster risk mitigation, and territorial land-use planning.

A critical technical challenge in constructing reliable landslide susceptibility models lies in the selection and assembly of non-landslide sample inventories. Non-landslide samples determine dataset quality and directly influence the predictive accuracy of trained classifiers. Common sampling strategies frequently introduce systematic biases: completely random selection across unfailed terrain can select points sharing identical engineering geological characteristics with actual failure zones; buffer zone sampling around known landslides depends heavily on subjective distance thresholds; and restricting sample selection to low-frequency zones or gentle slopes imposes strict spatial constraints that overlook regional spatial heterogeneity, increasing the vulnerability of classifiers to overfitting. To resolve these limitations, the research team implemented a stratified sampling approach rooted in engineering geological petrofabric classes, which categorizes bedrock based on joint density, lithological characteristics, and rock shear strength. By distributing non-landslide points proportionally across defined geotechnical intervals while enforcing strict minimum distance separations and buffer exclusions, this stratified methodology captures the complete spectrum of stable geological environments and minimizes spatial dependency across the terrain.

The study focused on the Maiji District in southeastern Tianshui City, an administrative territory spanning approximately 440 square kilometers within the southern Longxi Loess Plateau and bordering the western Qinling Mountains in the Weihe River Basin. The topography exhibits pronounced variation, descending from elevations of 2561 meters in the uplifted mountainous southeast to 755 meters in the northwest, where loess and red beds dominate. Tectonically situated along the active northern and southern segments of the Liupan Mountains seismic zone, the district contains diverse lithological exposures, including Quaternary alluvial and aeolian loess deposits, Neogene mudstones, and lower Paleozoic gneiss. A continental semi-humid monsoon climate concentrates seasonal rainfall between July and September, which readily infiltrates structurally weak geological contacts between loess mantles and bedrock. Field investigations and institutional records documented 436 geological hazard events across the district, comprising 362 landslides alongside collapses and debris flows, clustered heavily in northern loess hilly communities such as Mapaoquan, Huaniu, and Shetang.

To build a robust predictive database, researchers compiled a comprehensive spatial inventory containing the 362 confirmed landslide locations and an equal quantity of 362 non-landslide points selected across six petrofabric strata. Bedrock petrofabric divisions included river areas, hard rock, relative hard rock, relative soft rock, soft rock, and loose soil. Sampling excluded a 1000-meter buffer around known landslide sites, enforced a minimum 1.0-kilometer spacing between adjacent non-landslide coordinates to avoid spatial autocorrelation, and verified selected non-landslide locations against high-resolution remote sensing imagery. The combined sample collection was partitioned into a 70 percent training subset and a 30 percent independent validation subset. Concurrently, sixteen candidate conditioning indicators across topographical, geological, environmental, engineering, and triggering categories were standardized to a 30-meter raster grid resolution, deriving terrain indices from digital elevation data and environmental metrics from institutional lithological records, vegetation indices, and regional precipitation distributions.

Rather than relying on unvetted factor sets, the researchers implemented ablation experiments to filter redundant spatial predictors and enhance model stability across three boosting frameworks: Categorical Boosting, eXtreme Gradient Boosting, and Light Gradient Boosting Machine. By sequentially removing individual candidate factors and measuring the resulting decrement in the area under the receiver operating characteristic curve relative to full-factor baseline models, the ablation testing quantified each factor’s independent contribution. Factors whose removal caused notable performance drops—specifically elevation, slope gradient, engineering geological petrofabric, land use, and annual precipitation—were retained. Conversely, factors that provided negligible contributions or displayed informational redundancy, including surface roughness, terrain relief, topographic wetness index, and distance to faults, were systematically eliminated. This screening process established a finalized spatial database of twelve conditioning factors: elevation, slope, aspect, plan curvature, profile curvature, lithological petrofabric, distance to roads, distance to rivers, distance to earthquake epicenters, land use, normalized difference vegetation index, and annual rainfall.

The study subsequently deployed three advanced gradient boosting architectures to evaluate spatial susceptibility: Categorical Boosting, designed with symmetric tree topologies and specialized ordered target statistics to handle categorical variables without data leakage; eXtreme Gradient Boosting, incorporating regularization parameters and second-order loss function derivatives to curb overfitting; and Light Gradient Boosting Machine, utilizing histogram binning and leaf-wise tree growth with depth limits to maximize computational speed. When trained with default hyperparameter configurations, all three baseline algorithms exhibited noticeable signs of overfitting, achieving near-perfect training accuracies near 1.0 while test set accuracies plateaued between 77.31 percent and 78.78 percent. Among unoptimized baselines, Categorical Boosting yielded the top performance with an area under the curve of 0.844 and test accuracy of 78.78 percent, slightly surpassing Light Gradient Boosting Machine at 0.840 area under the curve and eXtreme Gradient Boosting at 0.834 area under the curve.

To mitigate overfitting and establish optimal structural parameters, the researchers applied three distinct hyperparameter optimization strategies to each boosting model: Random Search Optimization, Bayesian Optimization, and Particle Swarm Optimization. Random search sampled hyperparameter intervals stochastically, while Bayesian optimization employed probabilistic Gaussian processes and acquisition functions to explore the configuration space efficiently. Particle Swarm Optimization simulated collaborative flocking mechanics, updating particle velocities and positions through local and global memory to locate optimal parameter combinations. Across all combinations, hyperparameter tuning enhanced predictive accuracy, with Particle Swarm Optimization yielding the most pronounced gains across models. Specifically, Particle Swarm Optimization improved the Categorical Boosting model’s area under the curve by 3.4 percent and test accuracy by 2.4 percent. The resulting optimized particle-swarm Categorical Boosting framework achieved the highest comprehensive test performance among all twelve evaluated configurations, recording an area under the curve of 0.873, an overall accuracy of 80.6 percent, precision of 81.9 percent, recall of 81.9 percent, and an F1-score of 81.9 percent.

Applying the optimized Particle Swarm Optimization-tuned Categorical Boosting classifier to every 30-meter raster cell across Maiji District generated continuous landslide probability surfaces, which the researchers categorized into five susceptibility tiers—very low, low, moderate, high, and very high—using five statistical zoning methods. Comparing natural breaks (Jenks), equal intervals, geometric intervals, quantiles, and the mean-standard deviation technique revealed substantial zoning differences. Frequency ratio analysis, which tracks landslide event concentration relative to zonal land area, demonstrated that the natural breaks method achieved the most reliable spatial differentiation. Under natural breaks classification, the very high susceptibility tier covered 1.76 percent of the territory while achieving the highest frequency ratio of 3.93, whereas the very low susceptibility tier encompassed 45.39 percent of the land area. Conversely, geometric intervals allocated an excessive 35.73 percent of the district to the very high class, which proved practically implausible. Natural breaks appropriately delimited hazard clusters without distorting low-risk regional baselines.

The resulting susceptibility zoning delineated distinct spatial patterns governed by geomorphology and human intervention. Extensive central and southern mountainous zones were classified as very low and low susceptibility environments, characterized by elevated topography, undisturbed natural vegetation, minimal engineering development, and competent hard bedrock exhibiting high shear strength that resists sliding. In sharp contrast, high and very high susceptibility zones concentrated predominantly along the northwestern sector of the district, along the northeastern banks of the Weihe River, and beside its southeastern tributaries. These hazard-prone zones coincide with lower terrain altitudes, fragmented loess landforms, loose mechanical strata, and intense anthropogenic disturbance, where urban expansion, agricultural cultivation, and infrastructure excavations continuously disrupt slope configurations and accentuate underlying slope instability.

To overcome the black-box nature of ensemble tree algorithms, the authors integrated the Shapley Additive exPlanations framework, interpreting global factor importance and local single-variable dependencies using cooperative game theory principles. Feature summary rankings identified elevation as the primary controlling predictor, followed by slope gradient, engineering geological petrofabric, distance to roads, land use, and profile curvature, with secondary influences from river proximity, epicenter distance, precipitation, vegetation cover, aspect, and plan curvature. Single-factor dependency evaluations demonstrated that terrain elevations below 1500 meters generated positive Shapley values, driving landslide likelihood due to intense human engineering activities and unconsolidated loess deposits. Slopes under 25 degrees exhibited elevated historical failure frequencies because horizontal stratified structures and gentle inclines enhance rainwater retention in vertically jointed loess. Unconsolidated loose soils and soft rocks similarly contributed positive Shapley values due to low mechanical shear strength and susceptibility to weathering.

Shapley dependency plots further illustrated the decisive role of anthropogenic activity and slope morphology in destabilizing terrain. Distance-to-road values under 2.5 kilometers produced positive Shapley values, tracking how road excavation, artificial slope cutting, and mechanical construction vibrations disturb internal rock-mass stresses and induce failure conditions. In land-use classifications, croplands, engineering construction lands, and grasslands exhibited strong positive contributions to landslide occurrence. While deeply rooted forest and shrub vegetation mechanically reinforces surface soils and inhibits shallow sliding, agricultural irrigation and shallow-rooted grasses reduce stratum shear strength without stabilizing regolith. Additionally, positive profile curvature values between 0 and 15 indicated concave slope geometries, which act as natural topographic catchments, concentrating surface runoff, saturating pore spaces, and accelerating shear plane liquefaction.

The authors noted specific methodological limitations, emphasizing that empirical model precision remains constrained by the finite inventory of 362 historical landslide occurrences across the district. Limited disaster sample sizes restrict generalization, underscoring the necessity of continuously updated hazard databases derived from systematic field surveys. While algorithmic augmentation techniques such as Generative Adversarial Networks or Gaussian Mixture Models could synthetically expand minority samples, their influence on susceptibility modeling requires detailed future exploration. Additionally, while petrofabric-stratified sampling markedly stabilized boosting performance compared to conventional unconstrained selection, broader comparative analyses examining alternative stratified schemes remain an open research avenue. Overall, the integrated framework combining petrofabric stratification, particle swarm hyperparameter optimization, and game-theoretic interpretability establishes an empirical, verifiable blueprint for municipal hazard planning and precise geological disaster mitigation across dissected loess landscapes.

Subject of Research: Machine learning modeling for regional landslide susceptibility mapping

Article Title: Landslide susceptibility mapping based on stratified non-landslides sampling from engineering geological petrofabric and optimized boosting models: a case of Maiji District, China

Article References: Sang, K., Sun, P., Li, R., Ke, C., & Zhang, S. (2026). Landslide susceptibility mapping based on stratified non-landslides sampling from engineering geological petrofabric and optimized boosting models: a case of Maiji District, China. Geoenvironmental Disasters, 13(1), Article 38. https://doi.org/10.1186/s40677-026-00389-1

Image Credits: AI Generated

DOI: 10.1186/s40677-026-00389-1

Keywords: landslide susceptibility mapping, Maiji District, Categorical Boosting, Particle Swarm Optimization, engineering geological petrofabric, stratified sampling, Shapley Additive explanations, geological hazards, extreme gradient boosting, Light Gradient Boosting Machine, Bayesian optimization, random search optimization, receiver operating characteristic, area under the curve, Tianshui City, Weihe River Basin, slope stability