Abstract
Censored geochemical data, particularly below detection limits, challenge mineral exploration by biasing anomaly delineation and spatial patterns. This study presents a multi-stage framework combining Bayesian Gaussian Random Field (BGRF) modeling with Random Forest (RF) learning, enhanced by fractal-based background separation, to accurately predict censored Au concentrations. 14 samples with gold concentrations below 5 ppb were hypothesized as censored data to enable a more accurate evaluation of the model's performance based on their real Au concentrations. Unlike constant substitution methods, the framework preserves censored information and reconstructs spatial variability through probabilistic inference and nonlinear learning. The BGRF model incorporates spatial coordinates and Cu as the principal covariate to capture spatial autocorrelation and inter-element associations, producing probabilistic estimates for hypothesized censored data (HCD) that are then used to train the RF under a 5-fold out-of-fold scheme. The HCD estimated by spatial BGRF covariate model were performed as inputs for RF prediction model. A targeted calibration and scaling procedure reduces detection-limit bias and improves low-range predictions. Comparative analyses show that the calibrated and scaled RF-BGRF model substantially enhances accuracy and preserves realistic geochemical structures, outperforming half the detection limit (LD-half) or the detection limit divided by the square root of two (LD-rad2) approaches. This framework offers a promising tool for refining left-censored geochemical data in complex geological environments.