山崩
地形
地质学
特征(语言学)
职位(财务)
饱和(图论)
瞬态(计算机编程)
地图学
领域(数学)
碎片
遥感
边坡破坏
岩土工程
地貌学
水文学(农业)
钥匙(锁)
自然灾害
人工神经网络
作者
Junyu Wu,Qing Lü,Zhongxuan Liao,Yang Yu,Jie Song,Lei Wang
摘要
Abstract Rainfall‐induced shallow landslides often initiate near ridges, but why failure localizes at a particular upper‐slope position rather than elsewhere along the same hillslope remains unresolved. Terrain geometry, soil thickness, and transient pore‐water pressure are recognized controls, but how their spatially varying interactions along continuous slope profiles constrain initiation location remains insufficiently quantified. Here, we developed a deep‐learning‐based strategy to predict landslide initiation sites by analyzing continuous topography and soil thickness, capturing spatial dependencies ignored by using traditional cell‐based methods. Trained and validated on a landslide inventory from Zhuji, China, our approach achieves a mean normalized position error (MNPE) below 9%. Feature attribution ranked soil thickness, plan curvature, and slope angle as the three highest‐contributing predictors and showed that their contributions varied systematically along the profiles. To provide a process‐based interpretation of this localization pattern, we coupled a transient hydrologic model with a three‐dimensional slope‐stability model. The hydrologic model was calibrated using field observations from a monitored slope, and regional localization was evaluated against mapped initiation positions across 553 profiles. Modeled transient saturation exceeded the critical threshold at 85.0% of mapped initiation nodes, and the hydro‐mechanically estimated initiation positions yielded an MNPE of 14.9%. These results indicate that upper‐slope failures concentrate where rapid saturation development coincides with relatively low mechanical resistance, supporting a terrain‐conditioned hydro‐mechanical explanation for initiation in the Zhuji setting.
科研通智能强力驱动
Strongly Powered by AbleSci AI