WTCNN–Transformer: An unsupervised model for mineral prospectivity mapping based on wavelet transform, convolutional neural network, and transformer

• An unsupervised WTCNN–Transformer framework is developed for chromite prospectivity mapping. • Enhanced multi-source geoscience data through wavelet-based multiscale decomposition and reconstruction to improve anomaly characterization. • Incorporate multi-source geological, geochemical, and geophysical data to improve the accuracy of identifying high-mineralization potential areas. • The method achieves strong predictive performance (AUC = 0.8935) while reducing dependence on labeled samples. Strategic mineral resources are an important foundation for national resource security and industrial development. Traditional mineral prediction methods, constrained by their dependence on expert experience, limited surface information, and insufficient feature extraction capability, have difficulty in effectively identifying complex or concealed ore bodies. In this study, we propose a new unsupervised metallogenic prediction framework, WTCNN–Transformer, that integrates the wavelet transform, convolutional neural network, and Transformer. The core idea is as follows: the wavelet transform module performs multiscale decomposition and reconstruction of heterogeneous multi-source geoscience data, serving as a form of data enhancement and providing more discriminative inputs for subsequent feature extraction; the CNN module captures local spatial structural features; and the Transformer module models long range dependencies among regions through its self-attention mechanism. The framework adopts an unsupervised prediction strategy, automatically identifying intrinsic structures and metallogenic patterns in the geoscience data, which helps avoid problems caused by imbalanced positive and negative samples. Taking the chromite deposit area of Tuoli County in Xinjiang as an example, geophysical and geochemical data together with fault distribution and ophiolite bodies are used as key metallogenic controlling factors for joint training and prediction. The experimental results show that the framework achieves an AUC of 0.8935 and captures 83.3% of the known deposits within merely 16.6% of the predicted area. This performance is superior to that of the baseline models, providing theoretical support for chromite exploration in the region.