The core insight is deceptively simple: not all landslides fail the same way, so they should not be predicted the same way. In red-bed formations, interbedded siltstone, mudstone, marl, sandstone, and conglomerate are generally fractured and weakly cemented, and two dominant failure modes emerge. Retrogressive landslides, which account for 67 percent of the 164 documented events between 2004 and 2018 in the Sanshui inventory, begin at the slope toe when rising porewater pressure reduces shear strength and generates uplift forces that diminish sliding resistance, then retreat upslope in a progressive chain. Translational landslides, making up the remaining 33 percent, develop polygonal shear surfaces when rainwater infiltrates cracks at the trailing edge of mudstone slopes, increasing hydrostatic pressure and sliding stress while reducing effective normal stress on the potential failure plane. These are fundamentally different hydromechanical laws, and the researchers argue that folding them into a single statistical entity is precisely why traditional susceptibility maps fail in lithologically complex basins.

The framework unfolds in three stages. First, the team built a geographic information system database from multi-source data, including a 30-meter digital elevation model derived from a locally surveyed 1:10,000 contour map, Landsat 8 imagery for vegetation indices, meteorological records, land-use rasters, hydrographic and road networks, and geological maps. Eleven conditioning factors were derived: elevation, slope, aspect, plan curvature, profile curvature, NDVI, annual precipitation, land use, distance to river, distance to road, and lithology. The choice to generate terrain derivatives from the local contour map rather than public DEM products such as SRTM or ASTER GDEM matters, because low-relief red-bed basins hide subtle geomorphological variations that coarser global datasets smooth away, and the locally generated DEM preserved terrain details more effectively.

Second, the researchers quantified how strongly each factor interval relates to landslide occurrence using two strategies. The classical frequency ratio method computes the ratio of landslide density within a factor subinterval to the overall landslide density across the study area, a workhorse of bivariate susceptibility analysis. The Bayesian probability method goes further, integrating uncertainty and variability into the quantification by combining the conditional probability of a factor subclass given a landslide event with the prior probability of landsliding and the marginal probability of the subclass area. These quantified factor values then feed two machine learning classifiers, random forest and logistic regression, producing four hybrid models per landslide type: FR-RF, FR-LR, BP-RF, and BP-LR. Correlation screening confirmed that no pair of the eleven factors exceeded a coefficient of 0.5 under the Bayesian treatment, eliminating multicollinearity concerns before model training.

The results were unambiguous. Bayesian-enhanced models consistently outperformed their frequency-ratio counterparts for both landslide types, with the Bayesian random forest achieving the highest accuracy. For translational landslides, the BP-RF model yielded frequency ratios of 2.761 in high susceptibility zones and 6.635 in very high zones, beating FR-RF, FR-LR, and BP-LR. For retrogressive landslides, BP-based models increased frequency ratios in high and very high susceptibility categories by 4.12 to 60.41 percent compared with other methods, and the BP-RF model achieved the lowest very-low susceptibility frequency ratio of 0.037 alongside the highest very-high ratio of 4.920. The spatial patterns also made geological sense: high and very high susceptibility zones concentrated in the hilly southern region, where steep terrain, an extensive highway network, and converging rivers collectively raise landslide propensity, while the flatter northern and central basins mapped as low risk.

The third stage is where the study makes its most novel technical contribution. Rather than simply averaging the two type-specific susceptibility maps, the team introduced a mixture probabilistic neural network, or MPNN, that fuses the multi-modal susceptibility indices while respecting spatial heterogeneity. The architecture combines a mixture probabilistic model, grounded in Bayesian model averaging, with a spiking neural network stream built on back propagation. The mixture component computes weights for each landslide-type member through iterative updates of a Dirichlet distribution parameter, while the neural stream estimates geographically weighted parameters using a spatial distance matrix that relates the retrogressive susceptibility index at every raster cell to each translational landslide reference point. Unlike conventional spiking neural networks, the design does not model temporal spike propagation; instead, spatial dependence is represented through probabilistic weighting of neighboring landslide samples. A sigmoid output layer normalizes final susceptibility values between zero and one for each cell.

Performance evaluation against a unified model trained on a combined landslide inventory revealed the scale of the gain. The MPNN confined high and very high susceptibility zones to 16.42 percent of the total area yet captured 81.32 percent of all landslides, whereas the unified model designated 34.77 percent of the area as high to very high risk while containing only 76.09 percent of events. In very high susceptibility zones, the MPNN achieved a frequency ratio of 7.956, nearly triple the unified model’s 2.681. The overall AUC of 90.2 percent against 66.7 percent confirms that explicitly accounting for divergent failure mechanisms, rather than blending them, is what unlocks the accuracy. The team reserved 30 percent of the inventory strictly for independent validation, keeping all model construction procedures on the training subset.

To open the black box, the researchers applied SHAP, or Shapley additive explanations, a method rooted in cooperative game theory that quantifies each feature’s contribution to predictions. The analysis identified lithology, annual precipitation, and road proximity as dominant controls. For translational landslides, annual precipitation, road proximity, and lithology correlated positively with susceptibility, while elevation showed a negative correlation. For retrogressive failures, annual precipitation, lithology, and elevation correlated positively, while NDVI, a proxy for vegetation cover, mitigated risk. A Sankey diagram analysis of lithostratigraphic groups showed that Groups 1, 7, 12, and 13 exhibit the highest susceptibility, characterized by interbedded Cretaceous sandstone-mudstone sequences and vertically jointed siltstone layers. Group 12, a sand-mud interbed, destabilizes through permeability contrasts between sandy and muddy layers, while Group 13, siltstone-shale, provides weak shear surfaces under rainfall infiltration. These lithologies form anisotropic permeability systems that promote preferential groundwater flow, a preparatory factor for slope failure, further accentuated by gently inclined dip slopes with pervasive discontinuities where transient pore pressure rises during heavy rainfall can reduce effective stress to critical levels.

The implications extend well beyond one district. Red-bed basins underlie vast swaths of southern China, where rapid urbanization, climate change, and extreme weather have intensified landslide risk, and similar interbedded soft-rock terrains appear globally, from weathered crystalline slopes in southern Italy to progressive failures in sensitive clays of eastern Canada and Scandinavia. The framework’s lithostratigraphy-informed, mechanism-aware pathway offers a template for regional risk assessment that distinguishes fundamental geological predispositions from transient triggers, and it reduces the false classification of densely built-up urban zones as highly susceptible, a persistent flaw of unified models that overestimate risk in developed areas.

The authors are candid about limitations. Extrapolation to regions with distinct geological or climatic settings remains uncertain because the model relies on localized lithostratigraphic features. The framework uses static environmental factors rather than dynamic triggers such as rainfall intensity or piezometric monitoring, so it does not yet support real-time early warning, and although Bayesian elements are incorporated, predictive uncertainty from parameter variability and model structure is not fully quantified. Validation relied on a random training-testing split, which may still allow spatial dependence among neighboring landslides to produce optimistic estimates; the team proposes spatial block cross-validation and temporally independent validation as remedies. Future work includes integrating real-time monitoring data, automating landslide type classification through deep learning-based semantic segmentation of satellite imagery, and building hybrid data assimilation frameworks using UAV-LiDAR derived scarps. Even with these caveats, the message is clear: when the ground fails in more than one way, the map should know it.

Subject of Research: Earth Science

Article Title: Lithostratigraphically-constrained landslide susceptibility modeling in red-bed basins using a bayesian probability-based mixture probabilistic neural network

Article References: Liu, Z., Yi, A., Huang, G., Li, Q., & Zheng, T. (2026). Lithostratigraphically-constrained landslide susceptibility modeling in red-bed basins using a bayesian probability-based mixture probabilistic neural network. Frontiers in Earth Science, 14, Article 1910278. https://doi.org/10.3389/feart.2026.1910278

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1910278

Keywords: landslide susceptibility, red-bed basin, Bayesian probability, mixture probabilistic neural network, machine learning, Sanshui District, SHAP analysis, lithostratigraphy, Lithostratigraphically-constrained, landslide, susceptibility, modeling