Evaluation of a depth to basement Bayesian model using cross-validation strategies

Mapping the basement morphology undercover is crucial to identify structures significant for mineral exploration. Using geophysics, this mapping is done indirectly, and the solutions obtained present important uncertainty. Uncertainty can be estimated using Bayesian methods and can be reduced when introducing constraints such as direct drill hole observations of the depth to basement. Cross-validation techniques are commonly used to assess the predictive performance of machine learning or statistical methods. In our case, prediction applies to depth to basement and is observed using drill holes. However, a bias exists in the spatial distribution of the drill holes used for prediction. Certain exploration strategies, often based of geophysical observations, cause a dependency between adjacent observations that is incompatible with the statistical independence required by standard cross-validation techniques. This may cause overestimation of predictive performance. This paper compares different cross-validation strategies to assess their suitability to validating predictive performance of a workflow which maps depth to basement and its uncertainty in the Carrapateena province, South Australia, using Bayesian MT models, and drill holes measurements. Four validation strategies are considered:1) random k-folds cross-validation, 2) spatial block cross-validation, 3) depth block validation and 4) a random k-folds cross-validation with buffer. These results show that the spatial dependencies existing in the drill hole locations used for validation influences the estimation of the performance, and that a classic random k-folds cross-validation overestimates the performance that should be expected in application. We therefore recommend using random k-folds cross-validation with buffer or block validation strategies when using data derived from drill hole measurements for validation of predictive methods.