Title: Utilizing Local Intrinsic Dimensionality in Ground Motion Data to Anticipate Catastrophic Slope Failures

Abstract: The application of Local Intrinsic Dimensionality (LID) has demonstrated significant promise in identifying anomalies within high-dimensional datasets, particularly in detecting failures in granular media such as landslides. In this context, the timely and precise identification of failure zones is essential for implementing effective geohazard mitigation strategies. Nevertheless, this objective remains difficult to achieve due to the inherent spatial correlations and temporal dynamics embedded within surface displacement records. To bridge this gap, we introduce a novel unsupervised framework known as spatiotemporal LID (st-LID), which extends the traditional LID methodology to facilitate robust failure detection across landslide monitoring networks.

Our methodology is built upon three primary innovations: 1. Kinematic Enhancement: This feature integrates velocity metrics into the LID calculation, thereby capturing instantaneous deformation rates and short-term temporal fluctuations. 2. Bayesian Spatial Fusion: By employing Bayesian estimation to aggregate LID values across local spatial neighborhoods, this component embeds spatial correlations and mitigates the impact of localized noise. 3. Temporal Modeling (t-LID): This newly developed variant characterizes the long-term dynamics of displacement data, offering a sturdy temporal representation of displacement behavior.

Through the integration of these elements, st-LID is capable of identifying complex, multi-stage failure zones that are frequently missed by conventional methods. Comprehensive experimental results indicate that st-LID consistently surpasses current state-of-the-art unsupervised baselines in terms of both detection precision and lead time. Consequently, this approach establishes a reliable foundation for landslide early warning systems and targeted risk interventions, ultimately strengthening community resilience and preparedness strategies.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC