Unsupervised Autoencoder-Based Feature Ranking and Anomaly Detection for Porphyry Copper Prospectivity Mapping from Multi-Source Geospatial Datasets

The mineral system model formalizes the critical geological processes and mappable parameters that control ore formation, which can then be translated into spatial predictors used as input features in machine learning (ML)-based mineral prospectivity mapping (MPM). In most MPM studies, exploration evidence features are indeed derived from the mineral system model of the targeted deposit type. However, not all features produced in this way are necessarily informative or favorable for prospectivity analysis. This challenge can be addressed by using feature selection frameworks to identify the most relevant features before applying ML and deep learning (DL) algorithms for mathematical integration. To address this need, this study employs an unsupervised variational autoencoder (VAE) framework to evaluate and rank exploration evidence layers. The VAE quantifies feature importance through a systematic strategy that measures the sensitivity of reconstruction-error components, mean squared error (MSE), mean absolute error (MAE), and Kullback–Leibler (KL) divergence, to individual feature variations. In this way, the VAE ranks the exploration features and helps to identify those that are the most useful for prospectivity mapping. The proposed approach was applied to a real geo-dataset from a porphyry copper district in Iran. Based on the conceptual model of porphyry copper mineralization, 15 evidence layers were generated, including proximity to phyllic, argillic, propylitic, iron oxide, and silicification alteration zones; proximity to intrusive rocks, faults, and fault intersections; and geochemical maps of Cu, Mo, Sb, Pb, Zn, As, and W. The VAE-based ranking indicated that evidence layers related to hydrothermal alterations, intrusive rocks, and faults were the most influential exploration features, whereas geochemical evidence layers showed lower relative importance. Based on this evaluation, two modeling scenarios were considered: in the first, all available features were used, and in the second, only the features selected by the VAE framework were included. In both cases, the final prospectivity model was produced by an autoencoder (AE). For comparison, the prediction-area (P–A) plots of the two prospectivity models were generated using 14 known mineral occurrences as positive ground-truth labels, indicating that the model based on the selected features achieved a higher prediction rate (80%) than the model based on all features (72%). These results demonstrate that the evidence layers derived from the mineral system approach can benefit from unsupervised VAE-based evaluation, leading to improved performance of the prospectivity modeling.