Scholarlink

<p>The Duobaoshan area, located in the northeastern segment of the Central Asian Xing’an–Mongolia orogenic belt, represents a typical porphyry Cu–Mo metallogenic system. Both the Duobaoshan super-large and the Tongshan large-scale porphyry Cu–Mo deposits have been discovered in this region, indicating significant metallogenic potential. However, due to the complexity of geological structures and increasing burial depth, mineral exploration in the deep sections has become considerably more challenging. 3D mineral prospectivity mapping (MPM) can characterize the spatial pattern of mineralization prospectivity in 3D space and more comprehensively integrate depth information and spatial continuity, but in practical applications, it is often constrained by the dual challenges of scarce positive samples and lack of reliable negative samples, leading to instability in deep model training and insufficient generalization. To address this bottleneck, this paper proposes a “generative-weakly supervised” coupled 3D volumetric probabilistic prediction framework. This framework uses 3D ResUNet as the primary predictor, introduces a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) to learn the latent distribution from limited positive samples and generate high-confidence class-positive samples to alleviate sample scarcity, and adopts bagging-based positive–unlabeled learning (BPUL) to estimate the positive class prior and select reliable negatives, thereby constructing a training set closer to the true distribution. Based on the 3D multi-source geoscience dataset from the Duobaoshan ore district, which comprises geological, geophysical, drilling, and other data, this study compared four experimental configurations to evaluate the performance of different methods: (i) random negative sampling + 3D ResUNet (baseline); (ii) BPUL + 3D ResUNet (emphasizing the selection of reliable negatives); (iii) WGAN-GP + 3D ResUNet (emphasizing positive sample generative augmentation); and (iv) WGAN-GP + BPUL + 3D ResUNet (the coupled framework). These configurations were designed to assess the impact of different sampling and modeling strategies on the mineral exploration process. The results showed that the coupled framework outperformed each individual strategy in metrics such as AUC-ROC, F1 score, and precision. The exploration targets delineated based on this framework exhibited higher spatial coherence and volumetric connectivity. Overall, the positive-sample generative augmentation by WGAN-GP and the reliable negative identification by BPUL complement each other in 3D mineral prospectivity mapping. This combination improves the training stability and generalization of 3D ResUNet under conditions of extremely limited positive samples and unlabeled negative samples, providing a reusable modeling workflow for 3D exploration targeting in Duobaoshan and similar ore districts.</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

3D ResUNet with WGAN-GP Augmentation and Bagging-Based Positive–Unlabeled Learning for 3D Targeting, Duobaoshan District, Heilongjiang, China

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Xiumei Lv,
  • Gongwen Wang

摘要

The Duobaoshan area, located in the northeastern segment of the Central Asian Xing’an–Mongolia orogenic belt, represents a typical porphyry Cu–Mo metallogenic system. Both the Duobaoshan super-large and the Tongshan large-scale porphyry Cu–Mo deposits have been discovered in this region, indicating significant metallogenic potential. However, due to the complexity of geological structures and increasing burial depth, mineral exploration in the deep sections has become considerably more challenging. 3D mineral prospectivity mapping (MPM) can characterize the spatial pattern of mineralization prospectivity in 3D space and more comprehensively integrate depth information and spatial continuity, but in practical applications, it is often constrained by the dual challenges of scarce positive samples and lack of reliable negative samples, leading to instability in deep model training and insufficient generalization. To address this bottleneck, this paper proposes a “generative-weakly supervised” coupled 3D volumetric probabilistic prediction framework. This framework uses 3D ResUNet as the primary predictor, introduces a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) to learn the latent distribution from limited positive samples and generate high-confidence class-positive samples to alleviate sample scarcity, and adopts bagging-based positive–unlabeled learning (BPUL) to estimate the positive class prior and select reliable negatives, thereby constructing a training set closer to the true distribution. Based on the 3D multi-source geoscience dataset from the Duobaoshan ore district, which comprises geological, geophysical, drilling, and other data, this study compared four experimental configurations to evaluate the performance of different methods: (i) random negative sampling + 3D ResUNet (baseline); (ii) BPUL + 3D ResUNet (emphasizing the selection of reliable negatives); (iii) WGAN-GP + 3D ResUNet (emphasizing positive sample generative augmentation); and (iv) WGAN-GP + BPUL + 3D ResUNet (the coupled framework). These configurations were designed to assess the impact of different sampling and modeling strategies on the mineral exploration process. The results showed that the coupled framework outperformed each individual strategy in metrics such as AUC-ROC, F1 score, and precision. The exploration targets delineated based on this framework exhibited higher spatial coherence and volumetric connectivity. Overall, the positive-sample generative augmentation by WGAN-GP and the reliable negative identification by BPUL complement each other in 3D mineral prospectivity mapping. This combination improves the training stability and generalization of 3D ResUNet under conditions of extremely limited positive samples and unlabeled negative samples, providing a reusable modeling workflow for 3D exploration targeting in Duobaoshan and similar ore districts.