Geosci. Front. | 基于贝叶斯地统计融合机器学习与物理边坡稳定模型：一套考虑不确定性的滑坡预测框架

滑坡是极具破坏力的地貌灾害，滑坡易发性制图是区域减灾、国土规划的核心工具，当前主流建模分为数据驱动机器学习与物理机理模型两类，但二者均存在显著短板：随机森林等机器学习算法可高效捕捉滑坡与多源地形、岩土因子间复杂非线性关联，兼容多源异构空间数据，却缺乏岩土力学物理约束，面对极端降雨、罕见地形时易生成违背物理规律的预测结果，且难以量化预测不确定性；TRIGRS 等物理边坡稳定模型基于入渗理论与极限平衡原理模拟降雨诱发失稳过程，具备清晰物理解释，但高度依赖精细、全域分布的岩土水文参数，区域尺度常需采用地下水均质、土体性质分区简化等假设，模拟偏差较大，同时无法依托历史滑坡观测完成概率校准。现有融合方法多仅用物理模型筛选训练样本或向神经网络损失函数嵌入控制方程，深度模型还对样本量要求极高，缺少能够统一整合机器学习经验规律、边坡物理过程，同时显式量化空间预测不确定性的完整建模框架，针对热带风化山区、具备分层实测土壤剖面数据的一体化混合滑坡预测体系仍存在研究空白，为此本文开展相关集成建模研究。

Integrating machine learning and physics-based slope stability through Bayesian geostatistics: An uncertainty-aware framework for landslide prediction

A.L. Achu, Jobin Thomas, C.D. Aju, U. Surendran, Thomas Oommen, Girish Gopinath

https://doi.org/10.1016/j.gsf.2026.102402

Effective landslide susceptibility assessment requires a modelling framework that integrates heterogeneous geospatial information while explicitly accounting for predictive uncertainty. This study developed a Bayesian geostatistical approach that combined susceptibility prediction from a data-driven random forest (RF) classifier with factor of safety (Fs) output from a physics-based model — transient rainfall infiltration and grid-based regional slope-stability (TRIGRS) — using the integrated nested Laplace approximation (INLA). The framework produces calibrated and uncertainty-aware landslide susceptibility maps by integrating empirical relationships captured by RF, slope stability dynamics simulated by TRIGRS, and a spatially-structured residual variability represented through a latent Gaussian field. The RF model captures nonlinear associations between landslide occurrences and their conditioning factors, whereas TRIGRS encodes the physics of slope stability during rainfall infiltration. INLA couples these complementary signals and yields nonlinear posterior adjustments that suppress weak RF probabilities and strengthen mid-to-high values, while elevating low-to-mid ranges of the Fs distribution and tempering the upper tail. The resulting susceptibility surfaces are spatially smoother, better calibrated, and more interpretable than either input model. At the slope unit (SU) scale, the INLA model attained an area under the receiver operating characteristic curve (AUROC) of 0.96, outperforming RF (0.91) and TRIGRS (0.83). Spatial cross-validation confirmed consistently high discrimination across sub-basins. The posterior credible intervals provide spatially explicit uncertainty patterns, narrower in high-susceptibility hotspots (0.8–1) and wider at moderate probabilities (0.3–0.7), offering guidance for setting decision thresholds, prioritising field validation, and optimising monitoring network design. By combining machine learning (ML), physics-based modelling, and Bayesian spatial inference in a unified workflow, this scalable and transferable framework delivers interpretable, decision-relevant susceptibility estimates suitable for operational landslide risk management.