Understand changes at your site with confidence
Descartes Labs is proud to deliver anomaly detection algorithms for your GDS results that help identify potentially concerning movement over a site so that appropriate actions can be taken on the ground. This article summarizes the three types of anomalies available to help with your interpretation.
Overview
Results from GDS anomalies can contain millions of datapoints which can be overwhelming, making it hard to identify and focus on areas that may require more attention. Analyses of known failures that have occurred on mining infrastructure and natural features indicates that the signals in InSAR data that precede these events can be subtle and complex and commonly are not necessarily associated with the absolute total deformation of those points.
For that reason, Descartes Labs has developed an anomaly detection system that looks for statistically significant changes in the behavior of each point and outlines and highlight these areas into anomalies in three categories. Additionally, Iris gives you the ability to visualize alerts from previous analyses of the same site to help understand how these areas have moved or changed over time. Continue below to read more details on this methodology.
To calculate and view anomalies, activate the checkboxes within the processing toolbox in Iris.
Current Anomalies
Steady
Steady anomalies are based on the linear trend of each point over the past 120 days. Linear movement can indicate that an area is slumping continuously through time which can damage infrastructure, even without a catastrophic failure.
To be classified as a steady anomaly alert, points must have a linear slope above a threshold and an acceleration rate below a threshold. Additionally, the standard deviation of the linear fit is thresholded based on the velocity such that more tolerance is given to higher velocities. Simply, lower velocity points have a much stricter requirement to meet the steady anomaly classification, while higher velocity points allow a much more lenient fit.
Sudden
Sudden anomalies focus on more abrupt changes to a point's deformation history and for that reason the conditions to qualify as a sudden anomaly relate to both the acceleration and velocity of the point within the past 120s of monitoring.
To meet the requirements, the polynomial fit to the time series must have an acceleration term above a threshold as well as showing a statistically anomalous increase in velocity within the last 4 collects (48 days using Sentinel-1 data).
Complex
When both steady and sudden anomalies are flagged for a given point, the point is classified as having a complex anomaly. These points may warrant additional observation as their behavior shows both continuous movement with a significant increase in velocity over the most recent collections.
Historical Anomalies
Two additional features allow you to visualize historical anomalies over your site in reference to those calculated for the most recent collection period. Viewing historical anomalies can give better context as to the severity and persistence of an anomalous feature. It is common to find small isolated anomaly polygons that occur within a given analysis but are not present in the collections before or after. Conversely, if area is displaying anomalous deformation and the polygons associated with that feature are consistent and growing or shifting with subsequent analysis, it may be more cause to investigate.
Timeseries and Persistence Visualizations
Iris provides two ways to visualize historical anomalies. Both of these methods, anomaly polygons from the preceding two-thirds of the analysis window (240 days in the case of a 1 year Sentinel-1 analysis) will be overlain in the map window.
In the timeseries mode, polygons will be unfilled with the color of the outline corresponding to the date of the anomaly scaling from old (dark blue) to recent (yellow). This can help you determine if the historical anomalies still present a current issue of the movement has dissipated.
In persistence mode, the anomaly polygons are given an faint red opacity. Using this method, overlain polygons become increasing saturated as the overlapping opacities compound, meaning single occurrence anomalies will appear pale while persistent anomalies appear darker. Again, this can help determine the severity of an anomaly over time and help direct appropriate actions.
Anomaly Polygons
Anomalies are calculated on a point by point basis for each analysis, giving every point a state of either alerting (steady, sudden, complex) or not. For visualization, these points are polygonised and smoothed to give you the vector outline polygons available in Iris. Polygons also must meet minimum size thresholds and internal holes will be closed during the smoothing process. That means that it is possible to have non-anomalous points within an identified anomaly polygon where the majority of the points contained met the requirements to draw the shape.