Landslide early warning system: Single-centroid K-means and extreme value theory-based thresholds

Developing accurate landslide early warning systems (LEWS) for creeping landslides remains challenging, as fixed empirical thresholds have inherent limitations when processing non-stationary displacement data and may vary across regions. To address this issue, this study proposes an adaptive, unsupervised anomaly detection framework that integrates single-centroid K -means clustering and extreme value theory (EVT). Unlike conventional static approaches, this method establishes a dynamic reference frame to characterize normal creeping behavior, utilizing EVT to define probabilistic "Vigilance" and "Alarm" thresholds. By incorporating multivariate features such as displacement, velocity, acceleration, and cumulative rainfall, and utilizing principal component analysis (PCA) for visualization, the system robustly captures kinematic deviations. Its core innovation lies in the dynamic updating mechanism, which allows the cluster centroid and thresholds to evolve with incoming data, thereby adapting to baseline drift without manual intervention. Validation across seven real-world landslide cases demonstrates that the proposed system can effectively identify critical acceleration phases with minimal false alarms, thereby reducing reliance on subjective expert judgment. This proposed system offers an objective, lightweight, and interpretable solution for landslide risk mitigation and is readily operationalized for such practices.