10:45 〜 12:15
[STT39-P06] BINGHAM CANYON: Detecting Precursor Onset of Acceleration Using Temporal Saliency and Non-Parametric Spatial Clustering
キーワード：Inverse Velocity, Early Warning, InSAR, Precursor, Change Detection, Mining
Slope failures associated with mine landslides and other geotechnical land deformation events, often occur without much apparent warning, resulting in loss of life and property. In landslide-prone regions, satellite-based multi-temporal interferometric synthetic aperture radar (MT-InSAR) approaches, such as interferometric point target analysis (IPTA) is commonly used to study the spatio-temporal evolution of displacement velocities. For early warnings, displacement-based time of failure (ToF) forecasts such as 'inverse velocity' can be used. However, forecast accuracy depends upon the identification of the precursor date corresponding to the onset of acceleration (OoA) and is also hindered by the noise in MT-InSAR time series. This leads to a high false-alarm rate, something undesirable in an operational setting. Therefore, to improve the applicability of ToF in the active monitoring of mine slopes, a methodology for detecting salient precursors of ground instability should be developed that is robust to both temporal and spatial noise.
In this paper, we propose automated OoA saliency detection along with noise smoothing based on probabilistic models of time-series displacements as an antecedent to the inverse velocity technique to further enhance the operational applicability of ToF methods. This allows for advanced alerting of ground instability with a reduced false alarm rate by utilizing spatiotemporal land displacement measurements generated from the IPTA analysis. The saliency detection mechanism adopts fast Fourier transform (FFT) based spectral residual (SR), which serves as a compressed representation of the displacement sequence while identifying the anomalous part of the original sequence. It adapts the SR model from the visual saliency detection domain to time-series anomaly detection and has previously been used successfully for detecting anomalous disruption of web services.
Saliency-based detection is an online trend change detection mechanism to find a precursor date in the time-series displacement velocities at each point-target (also known as persistent scatterers) to establish an alert mechanism for future failures. Spatio-temporal smoothing based on probabilistic modeling of noise was performed to remove the noise of displacement velocities before the application of inverse velocity. This results in a subset of point targets and their ToF. Furthermore, clustering based on the density of points was applied to remove spatial outlier point targets, such as point targets that are spatially close but possess different ToF point targets. As opposed to the conventional parametric clustering algorithm ‘density-based spatial clustering of applications with noise’ (DBSCAN), an adaptive non-parametric method ‘spatial point event outlier detection’ (SPEOD) was employed. SPEOD employs a multi-level constrained Delaunay triangulation that automatically clusters points based on the distance between the points compared to the distance from points of other clusters. SPEOD works by creating a graph with edges of points using Delaunay triangulation and then these edges are further constrained (cut) on multiple levels. Based on the relative distance between the points we use SPEOD to cluster the points at each time and remove isolated outlier points.
In the paper, we demonstrate our methodology through a case study of the 2021 Bingham canyon tailings dam failure by analyzing the precursor OoA and predicted ToF based on the IPTA analysis of Sentinel-1 SAR time-series imagery.
In this paper, we propose automated OoA saliency detection along with noise smoothing based on probabilistic models of time-series displacements as an antecedent to the inverse velocity technique to further enhance the operational applicability of ToF methods. This allows for advanced alerting of ground instability with a reduced false alarm rate by utilizing spatiotemporal land displacement measurements generated from the IPTA analysis. The saliency detection mechanism adopts fast Fourier transform (FFT) based spectral residual (SR), which serves as a compressed representation of the displacement sequence while identifying the anomalous part of the original sequence. It adapts the SR model from the visual saliency detection domain to time-series anomaly detection and has previously been used successfully for detecting anomalous disruption of web services.
Saliency-based detection is an online trend change detection mechanism to find a precursor date in the time-series displacement velocities at each point-target (also known as persistent scatterers) to establish an alert mechanism for future failures. Spatio-temporal smoothing based on probabilistic modeling of noise was performed to remove the noise of displacement velocities before the application of inverse velocity. This results in a subset of point targets and their ToF. Furthermore, clustering based on the density of points was applied to remove spatial outlier point targets, such as point targets that are spatially close but possess different ToF point targets. As opposed to the conventional parametric clustering algorithm ‘density-based spatial clustering of applications with noise’ (DBSCAN), an adaptive non-parametric method ‘spatial point event outlier detection’ (SPEOD) was employed. SPEOD employs a multi-level constrained Delaunay triangulation that automatically clusters points based on the distance between the points compared to the distance from points of other clusters. SPEOD works by creating a graph with edges of points using Delaunay triangulation and then these edges are further constrained (cut) on multiple levels. Based on the relative distance between the points we use SPEOD to cluster the points at each time and remove isolated outlier points.
In the paper, we demonstrate our methodology through a case study of the 2021 Bingham canyon tailings dam failure by analyzing the precursor OoA and predicted ToF based on the IPTA analysis of Sentinel-1 SAR time-series imagery.