A novel copper prospectivity mapping framework using the AlphaEarth foundations and transformer

Traditional Mineral Prospectivity Mapping (MPM) relies heavily on expert-driven feature engineering and multi-source datasets, which are often costly and spatially inconsistent. Although remote sensing provides broad spatial coverage, conventional spectral features have limited ability to represent complex mineralization-related geological processes. In this study, we present AEF-Transformer, the first attempt to introduce AlphaEarth Foundations (AEF) semantic embeddings into MPM. The AEF model is pretrained on large-scale multi-modal Earth observation data and provides 64-dimensional, task-agnostic representations of geologically meaningful surface patterns. These embeddings are used as unified inputs and combined with a Transformer-based architecture to model long-range spatial dependencies and multi-scale mineralization structures. Experiments in the Tuwu copper belt, Eastern Tien Shan, China, show that the proposed framework outperforms CNN and MLP models, achieving a macro AUC exceeding 0.98 and producing spatially coherent prospectivity patterns consistent with known copper occurrences. Baseline experiments using raw Landsat-8 multispectral inputs further indicate that the improvement is mainly related to the semantic representation capability of AEF embeddings rather than the Transformer architecture alone. A percentile-based Confidence Index (CI) and majority-voting strategy were further introduced to mitigate model-related uncertainty and delineate high-confidence exploration targets. The proposed framework reduces reliance on handcrafted features and heterogeneous datasets, providing a robust, data-efficient, and uncertainty-aware approach for regional-scale mineral exploration.