Scholarlink

<p>Traditional deep learning algorithms (DLAs) were originally developed for domains like computer vision and natural language processing. However, without domain-specific architectural modifications and hyperparameter tuning tailored to geoscientific data constraints, DLAs struggle to capture intricate nonlinear relationships among ore-controlling factors in mineral prospectivity mapping (MPM). Besides, the deployment of DLAs in MPM is notably hindered by the scarcity of labeled samples and the intricate nature of hyperparameter optimization (HPO). To address these challenges, this paper introduces an innovative multi-scale feature extraction data augmentation (MSFE-DA) approach, coupled with an adaptive HPO framework for deep learning (AHOF-DL) that leverages Bayesian optimization algorithm for HPO. This methodology was successfully applied in MPM for Cu–Au polymetallic deposits in the Xuancheng ore district. The results indicated that the MSFE-DA approach increased the number of training samples by approximately five times through data augmentation, while preserving the geological spatial relationships among the samples. Simultaneously, models optimized by the AHOF-DL demonstrated statistically significant performance enhancements compared to unoptimized counterparts. Moreover, the proposed MSFE-DA method and AHOF-DL exhibit strong reusability and portability, rendering them applicable to diverse DL-based MPM tasks beyond this case study and providing a valuable reference for future DL-based research in geoscience.</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

Novel Adaptive Hyperparameter Optimization Framework and Multi-scale Feature Extraction Data Augmentation Method for Deep Learning-Based Mineral Prospectivity Mapping

  • 期刊论文

  • 发表日期：

    2025年12月25日

查看全文
  • Chaojie Zheng,
  • He Li,
  • Xiaohui Li,
  • Zhongliang Chen,
  • Rulin Zhang,
  • Feng Yuan

摘要

Traditional deep learning algorithms (DLAs) were originally developed for domains like computer vision and natural language processing. However, without domain-specific architectural modifications and hyperparameter tuning tailored to geoscientific data constraints, DLAs struggle to capture intricate nonlinear relationships among ore-controlling factors in mineral prospectivity mapping (MPM). Besides, the deployment of DLAs in MPM is notably hindered by the scarcity of labeled samples and the intricate nature of hyperparameter optimization (HPO). To address these challenges, this paper introduces an innovative multi-scale feature extraction data augmentation (MSFE-DA) approach, coupled with an adaptive HPO framework for deep learning (AHOF-DL) that leverages Bayesian optimization algorithm for HPO. This methodology was successfully applied in MPM for Cu–Au polymetallic deposits in the Xuancheng ore district. The results indicated that the MSFE-DA approach increased the number of training samples by approximately five times through data augmentation, while preserving the geological spatial relationships among the samples. Simultaneously, models optimized by the AHOF-DL demonstrated statistically significant performance enhancements compared to unoptimized counterparts. Moreover, the proposed MSFE-DA method and AHOF-DL exhibit strong reusability and portability, rendering them applicable to diverse DL-based MPM tasks beyond this case study and providing a valuable reference for future DL-based research in geoscience.