Early identification of an impending rockslide location via a spatially-aided Gaussian mixture model

Shuo Zhou, Howard D. Bondell, Antoinette Tordesillas, Benjamin I. P. Rubinstein, James BaileyThe Annals of Applied Statistics 2020 · DOI

Abstract

Movement of soil and rocks in an unstable slope under gravitational forces is an example of a complex system that is highly dynamic in space and time. A typical failure in such systems is a landslide. Fundamental studies of granular media failure combined with a complex network analysis of radar monitoring data show that distinct partitions emerge in the kinematic field in the early stages of the prefailure regime, and these patterns yield clues to the ultimate location of failure. In this study we address this partitioning of constituent units in complex systems by clustering the kinematic data, specifically, with a Gaussian mixture model. In addition, we assume that neighboring units should move together. As a result, spatial information is taken into account in our model so that spatial proximity is retained. Our case study of a rockslide from high resolution radar monitoring data shows that, by incorporating spatial information, our approach is more effective in revealing the dynamics of the system and detecting the location of a potential landslide, compared to the use of only the kinematics.

Introduction

Landslides and rockfalls represent significant geological hazards, yet they are rarely spontaneous events. Modern Slope Stability Radar (SSR) technology allows for the detection of surface movements with sub-millimeter precision, producing vast amounts of high-resolution data. However, the current industry standard for analyzing this data often relies on monitoring displacement rates at a few pre-selected locations, which may overlook the broader spatial dynamics of the slope.

Slope failure is a complex, "noisy" process. In the pre-failure regime, multiple areas may exhibit intermittent movement, a phenomenon known as strain localization. While some areas eventually stabilize, others accelerate toward collapse. Traditional clustering algorithms often struggle to distinguish between these transient movements and the actual failure zone because they do not account for the irregular, non-elliptical geometries characteristic of real-world landslides.

Methodology: The Spatially-aided GMM (sGMM)

To address these challenges, we developed the Spatially-aided Gaussian Mixture Model (sGMM). This model employs a "mixture of mixtures" approach to enforce spatial coherence—the physical principle that neighboring units of rock or soil are likely to move together.

  1. Kinematic Clustering: The model first uses a two-component Gaussian Mixture Model (GMM) to partition data into "active" (moving) and "stable" (stationary) regions based on displacement features.
  2. Spatial Embedding: The 2D coordinates of each sensor are incorporated into the feature space to ensure geographical relevance.
  3. Flexible Geometries: Standard GMMs typically assume elliptical clusters. To accommodate the complex, often "C-shaped" or irregular geometries of landslides, the sGMM nests additional GMM sub-components within each top-level cluster. This allows the model to represent complex spatial shapes by combining multiple smaller elliptical components.
  4. Optimization: Parameter estimation is handled via an Expectation-Maximization (EM) algorithm, while the Bayesian Information Criterion (BIC) is used to automatically determine the optimal number of spatial sub-components.

Experimental Evaluation

The model was evaluated using both synthetic and real-world datasets:

  • Synthetic Data: A set of 400 points with overlapping movement patterns across three distinct spatial shapes (SP1: simple block, SP2: checkerboard, SP3: arch) to test the algorithm's ability to recover ground truth.
  • Real-World Data: High-resolution radar data from an open-pit mine, consisting of 1,803 spatial locations monitored every six minutes over three weeks, leading up to a major rockfall.

Key Results and Discussion

Superior Pattern Recovery

In synthetic tests, the sGMM significantly outperformed baseline methods like K-Means and standard GMMs, particularly as the spatial complexity increased. While baseline methods struggled with non-circular shapes (SP2 and SP3), the sGMM maintained high fidelity to the original spatial patterns.

Comparison of clustering results across three synthetic spatial patterns (SP1-SP3). The sGMM column demonstrates the highest fidelity to the original ground truth patterns compared to K-Means and standard GMM variants.

Quantitative analysis using Normalized Mutual Information (NMI) scores confirms this visual evidence. As shown in the figure below, the sGMM maintains high accuracy across all spatial patterns, whereas the performance of traditional methods drops sharply for more complex geometries.

Normalized Mutual Information (NMI) scores for different clustering methods. The sGMM consistently achieves higher NMI, particularly for the complex SP2 and SP3 patterns.

Early Warning in Real-World Scenarios

In the real-world mine case study, the sGMM successfully identified the eventual failure region two days earlier than baseline methods. This 48-hour lead time is critical for safety protocols, allowing for the timely evacuation of personnel and equipment. The model achieved this by identifying "spatiotemporal persistence"—the consistent growth of an active cluster—while correctly filtering out transient, intermittent movements in other areas that baseline methods often flagged as false alarms.

Conclusion

By integrating spatial information directly into the clustering process, the sGMM provides a more robust and interpretable map of slope dynamics. The model's ability to handle complex geometries and filter kinematic noise makes it a powerful tool for early landslide detection. Future work will explore the integration of multimodal data, such as groundwater pressure and tiltmeter readings, to further enhance the predictive capabilities of the system.

Summary updated

Abstract

Movement of soil and rocks in an unstable slope under gravitational forces is an example of a complex system that is highly dynamic in space and time. A typical failure in such systems is a landslide. Fundamental studies of granular media failure combined with a complex network analysis of radar monitoring data show that distinct partitions emerge in the kinematic field in the early stages of the prefailure regime, and these patterns yield clues to the ultimate location of failure. In this study we address this partitioning of constituent units in complex systems by clustering the kinematic data, specifically, with a Gaussian mixture model. In addition, we assume that neighboring units should move together. As a result, spatial information is taken into account in our model so that spatial proximity is retained. Our case study of a rockslide from high resolution radar monitoring data shows that, by incorporating spatial information, our approach is more effective in revealing the dynamics of the system and detecting the location of a potential landslide, compared to the use of only the kinematics.
BibTeX
@article{zhou2020early,
  author = {Shuo Zhou and Howard D. Bondell and Antoinette Tordesillas and Benjamin I. P. Rubinstein and James Bailey},
  title = {{Early identification of an impending rockslide location via a spatially-aided Gaussian mixture model}},
  journal = {The Annals of Applied Statistics},
  year = {2020},
  doi = {10.1214/20-aoas1326},
}
Figure