Scholarlink

<p>As machine learning gains widespread application in regional mineral prospectivity prediction, multi-source mineralization data commonly suffer from sparse sampling and uneven spatial distribution, imposing higher demands on prediction accuracy and stability. Traditional supervised learning struggles to simultaneously address geological spatial consistency and multidimensional attribute similarity and heavily relies on large-scale labeled samples. However, mineralization, as a rare geological phenomenon, often results in insufficient valid labels. To address this, this paper proposes a spatial–attribute graph adaptive model (SAGAM), a self-supervised learning framework with supervised fine-tuning. The model adaptively constructs a graph that integrates spatial proximity and geochemical feature similarity, enabling joint representation of spatial–attribute information. During the pretraining stage, a graph masked autoencoder (GraphMAE) is employed to reconstruct masked node features, allowing the model to learn global representations and structural dependencies from unlabeled data. During the fine-tuning stage, the pretrained model is optimized on a limited number of labeled samples through a supervised mineralization potential classification task. Taking Laos as the study area, experimental results demonstrate SAGAM's superiority over supervised graph convolutional networks (GCNs) and attention-based graph attention networks (GATs) across all metrics. Its high-probability mineralization zones exhibit stronger spatial concordance with known deposits and geochemical anomalies, significantly enhancing feature discriminability under limited samples. This provides an intelligent modeling framework for regional mineral prospectivity mapping.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SAGAM: A Self-Supervised Graph Learning Framework with Supervised Fine-Tuning for Mineral Prospectivity Mapping

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Miao Xie,
  • Bimin Zhang,
  • Xueqiu Wang,
  • Jian Zhou,
  • Wei Wang,
  • Hanliang Liu,
  • Dongsheng Liu,
  • Sounthone Laolo,
  • Phomsylalai Souksan,
  • Chunfang Dong,
  • Haonan Wang,
  • Zimu Yang

摘要

As machine learning gains widespread application in regional mineral prospectivity prediction, multi-source mineralization data commonly suffer from sparse sampling and uneven spatial distribution, imposing higher demands on prediction accuracy and stability. Traditional supervised learning struggles to simultaneously address geological spatial consistency and multidimensional attribute similarity and heavily relies on large-scale labeled samples. However, mineralization, as a rare geological phenomenon, often results in insufficient valid labels. To address this, this paper proposes a spatial–attribute graph adaptive model (SAGAM), a self-supervised learning framework with supervised fine-tuning. The model adaptively constructs a graph that integrates spatial proximity and geochemical feature similarity, enabling joint representation of spatial–attribute information. During the pretraining stage, a graph masked autoencoder (GraphMAE) is employed to reconstruct masked node features, allowing the model to learn global representations and structural dependencies from unlabeled data. During the fine-tuning stage, the pretrained model is optimized on a limited number of labeled samples through a supervised mineralization potential classification task. Taking Laos as the study area, experimental results demonstrate SAGAM's superiority over supervised graph convolutional networks (GCNs) and attention-based graph attention networks (GATs) across all metrics. Its high-probability mineralization zones exhibit stronger spatial concordance with known deposits and geochemical anomalies, significantly enhancing feature discriminability under limited samples. This provides an intelligent modeling framework for regional mineral prospectivity mapping.