Spatial Ensemble Anomaly Detection Method for Exhaustive Map-Based Datasets

Abstract

Spatial anomaly detection is an essential part of subsurface data quality control and modeling for resources, e.g., groundwater aquifer, hydrocarbon resources, and mining mineral grades, along with environmental remediation. Efficiently identifying spatial local anomalous regions is important for detecting changes in subsurface population and data acquisition artifacts. Advanced data analytics and machine learning methods may be applied, but often omit essential spatial context with the additional cost of reduced interpretability. Considering the high cost of the risk in subsurface resource exploration, it is essential to maximize the integration of domain expertise. Our proposed method is an ensemble anomaly detection method that calculates the local anomaly probability based on the joint probability density space derived from the semivariogram model. The ensemble approach of our proposed method utilizes multiple local anomaly classifications calculated over a moving search window around each grid of a 2D map or 3D model. The anomalous regions are eventually decided based on the majority rule of the ensemble anomaly classifiers. We demonstrate the proposed method with 3 synthetic exhaustive, map-based, spatial datasets to cover different scenarios for spatial anomaly applications. The proposed ensemble anomaly detection method integrates domain expertise, spatial continuity, and scale of interest. We suggest using the proposed method as an automated tool to effectively identify the spatial local anomalies, based on the rigorous use of spatial continuity and volume variance from geostatistics, to focus professional time.