The successful application of geographic information system (GIS)-based mineral prospectivity mapping (MPM) essentially relies on two factors: one is reasonable evidential layers that conform to geological cognition, and the other is excellent models that can extract critical prospecting information from evidential layers. Geological features in MPM are usually discretized by categorizing them into classes and assigning the same or linear weights to each class, which suffer from bias in the interpretation of geological processes under ambiguous knowledge. Moreover, either unsupervised or supervised MPM models are constructed based on the assumption that the variables are relatively independent or identically distributed. In terms of these two issues, this study develops a joint workflow that combines a rational evidence layer weighting method and a deep learning MPM model, considering both the spatial and genetic associations of geological features. A data-driven singularity-based weighting method is first applied to evaluate the relative importance of geological features for mineralization and assign continuous weights to evidential layers using a nonlinear function that is consistent with existing geological models. Then, a more recent deep learning model, namely the long short-term memory network, is employed to extract and integrate the deep-level geological prospecting information among the weighted evidence layers. This joint approach was demonstrated with the help of a case study targeting Fe mineralization in southwestern Fujian Province, China. The mineral potential map obtained using this approach revealed that almost all the known Fe mineral deposits developed in the delineated high prospective regions, indicating that the proposed workflow is reasonable and meaningful for MPM.
Read full abstract