Combining sequential Gaussian co-simulation and Monte Carlo dropout-based deep learning models for geochemical anomaly detection and uncertainty assessment

Geochemical anomaly detection is crucial for guiding mineral exploration toward prospective mineral deposits. However, this task is inherently challenging due to uncertainties arising from sparse sampling, spatial variability of geochemical patterns, model limitations. To evaluate the uncertainties associated with spatial variability and model limitations, this study proposes an innovative approach that combines sequential Gaussian co-simulation (SGCS) with Monte Carlo (MC) Dropout-based Convolutional Neural Networks (CNNs) for geochemical anomaly detection and uncertainty quantification. The SGCS method generates multiple realizations of geochemical data, facilitating the quantification of uncertainty in geochemical patterns by considering potential distributions at unsampled locations. These realizations are utilized to augment the training dataset for CNNs, thereby enhancing the model's robustness in anomaly detection. The MC Dropout technique is integrated into the CNN model to evaluate prediction uncertainties, providing critical insights for decision-making under uncertainty. The proposed methodology was applied to the northwestern part of Sichuan Province, China, a region known for gold mineralization. Results indicate that all known gold deposits fall within areas where the anomaly probability exceeds 0.843. By integrating predicted probabilities with associated uncertainties, the spatial distribution of a confidence index is derived, offering a structured guide for subsequent exploration. This integrated framework enhances anomaly detection accuracy and provides robust uncertainty estimates, ultimately enabling more efficient and informed exploration strategies in high-uncertainty environments.