A Self-Supervised Learning Framework Integrating Masked Autoencoders and Low-Rank Adaptation for Mineral Prospectivity Mapping in Western Shandong Province

Transport in Porous Media · Springer Nature · 2026-09-23 · DOI: 10.1007/s11004-026-10345-2

Аннотация

Mineral prospectivity mapping encounters difficulties in deriving reliable indicators from diverse geoscience data, especially when labeled data is extremely limited. To address this, we proposed a two-stage self-supervised learning framework that integrates masked autoencoders (MAE) and low-rank adaptation (LoRA) for skarn Au–Cu mineral prospectivity mapping in the Yinan-Yiyuan Area, western Shandong Province, China. We first utilized MAE to pretrain a vision transformer (ViT) on unlabeled multi-source data, enabling the learning of general spatial representations through masked reconstruction. Subsequently, LoRA was used to fine-tune the model with a limited number of mineral deposits (occurrences), constraining parameter updates to prevent overfitting. The results demonstrated that our proposed method significantly outperforms convolutional neural networks (CNN) and supervised ViT baselines, achieving area under the receiver operating characteristic curve (ROC-AUC) and precision–recall curve (PR-AUC) scores of 0.95 and 0.97, respectively. These results demonstrated that the self-supervised learning framework can effectively leverage abundant unlabeled geoscience data and improve mineral prospectivity mapping when labeled deposit data are scarce.

Читать у издателя →

Похожие статьи