Abstract Lithological mapping is essential for the exploration of critical minerals supporting energy transition and national defense. Although recent advancements have incorporated multi‐source data sets and leveraged machine learning and deep learning (DL) methods, lithological mapping continues to face significant challenges, such as data imbalance, limited availability of data, and the presence of mislabeled observations. Previously proposed approaches to address these limitations may introduce additional errors and become sources of uncertainty by altering the prior or lacking appropriate marginalization, thereby underscoring the need for uncertainty quantification. In this study, we formulate lithological mapping as a Bayesian inference problem within an image‐to‐image translation framework and propose a Swin Transformer‐based DL model to approximate the posterior probability distribution of lithological models. For applications in two regions under Phanerozoic cover in Canada, the Hudson Bay Lowlands and southwestern Manitoba, we utilize 48 features extracted from aeromagnetic and gravity data sets and generate labels using the most recent geological map of Canada. The proposed model exhibits robustness to data imbalance in the training data set and achieves training accuracies of up to and prediction accuracies of up to . Furthermore, we present a theoretical analysis of key sources of uncertainty, including data imbalance correction strategies, data augmentation via geophysical forward modeling, and label inaccuracies.
Ding et al. (Thu,) studied this question.