Scholarlink

<p>Traditional approaches to lithology mapping typically call for significant geological expertise and hands-on field experience, creating barriers for those without such expertise. In the past two decades, remote sensing technology has become an indispensable tool for lithological mapping. Recent years have witnessed a paradigm shift in this domain brought about by the integration of deep learning algorithms (DLAs). DLAs, notably convolutional neural networks (CNNs), have become prominent in geological applications. This research aimed to generate a lithological map of the Pariz-Chahargonbad area, renowned for its significant metallic mineral deposits. To do so, we capitalized on the combined capabilities of ASTER and Sentinel-2 data to identify and map lithological units. A sub-pixel technique involving multiple sequential steps was employed to extract consistent and desirable training data, which was then fed into CNN. Although CNNs wield their formidable power and utility through their ability to produce highly accurate predictions, applying these models to lithological mapping poses several challenges. To bolster the effectiveness of CNNs, ResNet-18 and GoogleNet pre-trained networks were leveraged through transfer learning, and the Harris hawks optimizer (HHO) was used to fine-tune the CNN parameters. As determined from the confusion matrices (CMs), the ResNet-18 CNN model demonstrated an accuracy of 84.6%. In a comparative evaluation, it became apparent that the GoogleNet CNN model outperformed the ResNet-18 CNN model, exhibiting a superior accuracy of 87.8%. The integration of HHO further improved the performance of the models. The accuracies of the HHO-ResNet-18 and HHO-GoogleNet CNNs were 97.6 and 99.33%, respectively. The achieved level of accuracy was confirmed by receiver operating characteristic diagrams, Recall, F1-score, and precision. Besides, a field validation campaign in the Kuh Panj and Bagh Khoshk regions was conducted to assess the accuracy of lithological classifications, confirming a high degree of concordance between field evidence and model predictions. The synergy between pre-trained CNN models and HHO optimization streamlined the identification and characterization of lithological units and significantly improved model accuracy and efficiency. These models'&#xa0;remarkable accuracy and adaptability are promising for various applications, including geological studies, mineral exploration, and environmental monitoring, potentially underpinning informed decision-making and promoting sustainable resource management.</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

Precise Lithological Mapping Using HHO-Optimized Deep ResNet-18 and GoogleNet CNN Models

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Yousef Bahrami,
  • Hossein Hassani,
  • Abbas Maghsoudi

摘要

Traditional approaches to lithology mapping typically call for significant geological expertise and hands-on field experience, creating barriers for those without such expertise. In the past two decades, remote sensing technology has become an indispensable tool for lithological mapping. Recent years have witnessed a paradigm shift in this domain brought about by the integration of deep learning algorithms (DLAs). DLAs, notably convolutional neural networks (CNNs), have become prominent in geological applications. This research aimed to generate a lithological map of the Pariz-Chahargonbad area, renowned for its significant metallic mineral deposits. To do so, we capitalized on the combined capabilities of ASTER and Sentinel-2 data to identify and map lithological units. A sub-pixel technique involving multiple sequential steps was employed to extract consistent and desirable training data, which was then fed into CNN. Although CNNs wield their formidable power and utility through their ability to produce highly accurate predictions, applying these models to lithological mapping poses several challenges. To bolster the effectiveness of CNNs, ResNet-18 and GoogleNet pre-trained networks were leveraged through transfer learning, and the Harris hawks optimizer (HHO) was used to fine-tune the CNN parameters. As determined from the confusion matrices (CMs), the ResNet-18 CNN model demonstrated an accuracy of 84.6%. In a comparative evaluation, it became apparent that the GoogleNet CNN model outperformed the ResNet-18 CNN model, exhibiting a superior accuracy of 87.8%. The integration of HHO further improved the performance of the models. The accuracies of the HHO-ResNet-18 and HHO-GoogleNet CNNs were 97.6 and 99.33%, respectively. The achieved level of accuracy was confirmed by receiver operating characteristic diagrams, Recall, F1-score, and precision. Besides, a field validation campaign in the Kuh Panj and Bagh Khoshk regions was conducted to assess the accuracy of lithological classifications, confirming a high degree of concordance between field evidence and model predictions. The synergy between pre-trained CNN models and HHO optimization streamlined the identification and characterization of lithological units and significantly improved model accuracy and efficiency. These models' remarkable accuracy and adaptability are promising for various applications, including geological studies, mineral exploration, and environmental monitoring, potentially underpinning informed decision-making and promoting sustainable resource management.