On the classification and treatment of outliers in a spatial context: A Bayesian Updating approach

AUTOR(ES)
FONTE

REM, Int. Eng. J.

DATA DE PUBLICAÇÃO

2021-09

RESUMO

Abstract Checking and treating extreme values is commonplace in modelling workflows. The main methods to manage outliers may be categorized into graphical, Kriging- and simulation-based approaches. While graphical methods usually classify outliers from a global perspective, geostatistical methods evaluate outliers in a local context. Ordinary-Kriging based approaches are affected by conditional bias associated with the distribution tail(s), impacting on the correct classification of extreme values; the simulation method is based on the fact that geostatistical simulation is robust for outlier values. However, this approach ignores the interaction among outliers in the same neighborhood. The proposed approach considers that there are two values available at every sampled position, the sampled value and the conditional probability estimated from nearby data through cross-validation; the sampled value. Each value outside the user-defined threshold is classified as an outlier and is edited by merging the sampled and kriged value through Bayesian Updating. The proposed method is performed in normal-score units using Simple Kriging to (i) correctly estimate conditional distributions in the cross-validation step; (ii) avoid conditional bias; and (iii) minimize the outlier influence on experimental-variogram modelling. The proposed method is compared to three other widely used methods in a case study of a gold deposit. The proposed method substantially improved the local accuracy and reduced the number of misclassified blocks of a reference model.

Documentos Relacionados