Geologically Constrained Convolutional Neural Network for Mineral Prospectivity Mapping

Various deep learning algorithms (DLAs) have been successfully employed for mineral prospectivity mapping (MPM) to support mineral exploration, due to their superior nonlinear extraction capabilities. DLAs algorithms are typically purely data-driven approaches that may ignore the geological domain knowledge. This renders the predictive results inconsistent with the mineralization mechanism and results in poor interpretation. In this study, a geologically constrained convolutional neural network (CNN) that involves soft and hard geological constraints was proposed for mapping gold polymetallic mineralization potential in western Henan Province of China. A penalty term based on the controlling equation of the spatial coupling relationship between the ore-controlling strata and gold deposits was constructed as a soft constraint to guide the CNN model training according to additional prior geological knowledge. In addition, domain knowledge related to mineralization processes and a geochemical indicator were simultaneously embedded as hard constraints in the feature extractor and classifier of the CNN, respectively, to control the model training based on the mineralization mechanism. The comparative experiments demonstrated that the geologically constrained CNN was superior to other models, thus indicating that the coupling of data and domain knowledge is effective for MPM and further improves the rationality and interpretability of the obtained results.