A multi-method explainability framework applied to machine learning predictions of landslide susceptibility has revealed how spatial mapping units alter the internal reasoning of random forest algorithms, demonstrating that slope units align more closely with terrain-driven physical mechanisms than conventional geometric grid units. In an investigation focused on Pingyang County in southeastern Zhejiang Province, China, researchers evaluated two widely used spatial units across identical environmental factor inputs and historical failure inventories. While both models demonstrated strong predictive accuracy, reaching area under the receiver operating characteristic curve scores above 0.9, deeper examination through model-agnostic interpretability metrics uncovered marked differences in how each unit structure captures environmental triggers, topographic relationships, and site-specific failure probabilities.
Landslide susceptibility prediction serves as a primary foundation for delineating hazardous terrain, informing land-use planning, and mitigating slope failure risks, particularly across mountainous territories where undetected ground instability causes severe human casualties and economic losses. While decision-tree ensemble methods such as random forests have become standard predictive tools due to their robustness, capacity to process high-dimensional spatial datasets, and resilience against overfitting, their operation as complex statistical black boxes has limited scientific acceptance. Analysts frequently evaluate models solely through classification accuracy metrics, leaving the underlying decision pathways unexamined. Concurrently, practitioners must choose spatial mapping units to subdivide landscapes, typically deciding between computationally efficient grid units or terrain-conforming slope units, without a systematic understanding of how this spatial discretization influences algorithmic logic.
To overcome the diagnostic constraints of solitary interpretability metrics, the investigative team formulated an integrated post-hoc analytical structure bridging global factor contributions down to individual slope failures. The analytical pipeline incorporates mean squared error feature importance ranking to isolate dominant conditioning variables, partial dependence plots and individual conditional expectation curves to track marginal responses across individual factor ranges, and two-dimensional accumulated local effects plots to quantify non-linear interactions between paired features even in the presence of correlated data. The framework concludes with full-feature Shapley additive explanations value decomposition applied to actual ground failures, enabling field-verified cross-checks between algorithmic attributions and physical geological site surveys.
The experimental protocol targeted Pingyang County, an area spanning roughly 1051 square kilometers of land dominated by complex mountainous and hilly geomorphology comprising 56.1 percent of the territory, alongside low-lying coastal plains covering 43.9 percent. The regional landslide inventory documented 225 historical failure surfaces recorded between 2000 and 2018, primarily comprising small-scale shallow soil slides with an average footprint of 1070 square meters concentrated at slope toes and mid-slope sections. These failures typically occur within residual clayey cohesive soils mixed with gravel overlying basal bedrock, triggered by concentrated precipitation during the July to September monsoon season as well as artificial toe excavation. The investigation established a 16-variable environmental factor database spanning topography, lithology, hydrology, and surface cover, ensuring that pairwise Pearson correlation coefficients remained below 0.7 to avoid severe multicollinearity.
Spatial discretization established the primary comparative variable between the models. For the grid unit configuration, the territory was partitioned into regular 8-meter raster cells, generating 3311 landslide cells from the 225 failure footprints. For the slope unit design, natural slope segments were extracted via multi-scale segmentation guided by topographic curvature and hydrological flow attributes, producing 225 individual landslide units. Continuous attributes were averaged across slope units while categorical features were assigned based on zonal mode values. To maintain parity and counteract class imbalance, equal numbers of non-landslide samples were selected outside flat plains, creating balanced one-to-one positive and negative datasets split into training and testing partitions at a seven-to-three ratio. Optimal random forest hyperparameters were tuned using Python and scikit-learn, configuring 500 estimators and tree depths of 15 for both models, with split and leaf parameters adjusted to suit the unit scales.
Standard accuracy assessments revealed closely matched global performance between the two discretization methods. The grid unit random forest achieved an area under the receiver operating characteristic curve of 0.964 and a prediction accuracy rate area of 0.947, whereas the slope unit configuration yielded values of 0.955 and 0.936, respectively. Furthermore, examination of landslide susceptibility index distributions showed that grid units produced an average index of 0.234 with a standard deviation of 0.216, compared to an average of 0.331 and a standard deviation of 0.245 for slope units. While both distributions aligned with valid predictive behavior where lower indices concentrate across stable plains and dispersed indices avoid overconcentration, standard statistical evaluation metrics alone failed to demonstrate which spatial framework generated more physically grounded assessments.
Deploying the integrated interpretability pipeline exposed the distinct mechanisms driving each model’s predictions. Global mean squared error feature importance identified slope angle and the normalized difference built-up index as the primary governing factors across both mapping units, while soil composition and soil type exhibited minimal predictive influence. However, slope units placed substantially higher importance on the topographic wetness index and the normalized difference vegetation index than grid units. Partial dependence and individual conditional expectation curves further detailed this behavior: susceptibility rose sharply as terrain inclination transitioned from 7.5 to 10 degrees, leveled off above 10 degrees, and began declining once slope angles reached 25 degrees. This decline directly reflects geotechnical conditions in the region, where steep gradients exceeding 25 to 35 degrees inhibit the accumulation of unconsolidated colluvial soil mantles required to initiate shallow landslides.
Two-dimensional accumulated local effects analysis revealed notable differences in multi-factor interactions between the two spatial frameworks. Examining the joint impact of slope angle and built-up indices demonstrated that grid units experienced sharp suppression of landslide probability when slope angles exceeded 35 degrees while built-up indices remained below 0.3, whereas slope units showed maximum suppression on gentle slopes under 10 degrees with built-up indices between 0.3 and 0.4. When evaluating slope angle alongside elevation, both units identified terrain below 10 degrees and above 600 meters as strongly conducive to failure, but slope units displayed substantially greater volatility in accumulated local effects values, showing marked sensitivity to subtle environmental shifts across low-elevation foothills between 15 and 20 degrees.
At the local validation scale, the researchers examined a field-verified shallow failure in Yuexi Village that occurred on August 8, 2015, following intense rainfall from Typhoon Sudirman, which caused structural damage to two houses. The failure involved a one- to two-meter layer of residual colluvial clayey soil resting on heavily weathered bedrock along a natural slope of 20 to 25 degrees beneath an upper concave terrain gully. Projecting model outputs onto three-dimensional remote sensing surfaces confirmed that both architectures classified the site as highly susceptible. However, full-feature Shapley additive explanations attributions demonstrated that the slope unit model assigned larger positive contribution values to critical terrain, vegetation, and human activity variables, resulting in a higher net susceptibility index that aligned closely with the observed water-focusing geometry of the slope.
The study highlights practical operational guidelines alongside explicit methodological caveats. Slope units prove superior when geological investigation requires strict fidelity to geomorphological processes, hydrologic catchment dynamics, and individual slope mechanisms, provided high-resolution environmental datasets are available to withstand their heightened data sensitivity. Conversely, regular grid units offer robust and computationally streamlined options for regional reconnaissance when source data resolutions are coarse. Methodological constraints persist, including reliance on coarse national datasets for rainfall and soil characteristics, random sampling limitations among non-landslide points, and the loss of internal topographic variation when averaging continuous variables within large slope units, emphasizing the need for higher-resolution remote sensing and automated geological verification in future hazard mapping.
Subject of Research: Explainable machine learning and mapping unit selection in landslide susceptibility prediction
Article Title: Mapping unit selection in landslide susceptibility prediction: A comprehensive explainable framework integrating multi-method interpretability
Article References: Huang, F., Jiang, B., Chen, Z., Rezania, M., Chang, Z., Li, H., & Zhang, Z. (2026). Mapping unit selection in landslide susceptibility prediction: A comprehensive explainable framework integrating multi-method interpretability. International Journal of Coal Science & Technology, 13(1), Article 62. https://doi.org/10.1007/s40789-026-00897-3
Image Credits: AI Generated
DOI: 10.1007/s40789-026-00897-3
Keywords: landslide susceptibility prediction, explainable artificial intelligence, random forest, mapping units, slope units, grid units, partial dependence plots, individual conditional expectation, accumulated local effects, shapley additive explanations, Pingyang County, geological hazards, topographic wetness index, feature importance, receiver operating characteristic
Cite Scienmag News
Blake Davidson. (November 21, 2026). Explainable Framework Clarifies How Mapping Units Influence Landslide Prediction Models. Scienmag. https://scienmag.com/explainable-framework-clarifies-how-mapping-units-influence-landslide-prediction-models/
Blake Davidson. "Explainable Framework Clarifies How Mapping Units Influence Landslide Prediction Models." Scienmag, 21 November 2026, https://scienmag.com/explainable-framework-clarifies-how-mapping-units-influence-landslide-prediction-models/. Accessed 21 November 2026.
Blake Davidson. "Explainable Framework Clarifies How Mapping Units Influence Landslide Prediction Models." Scienmag. November 21, 2026. https://scienmag.com/explainable-framework-clarifies-how-mapping-units-influence-landslide-prediction-models/








