Scholarlink

<p>The automatic detection of geological lineaments from satellite imagery remains a challenge due to the limitations of traditional methods and the complexity of both natural and anthropogenic structures. This study proposes an integrated approach combining convolutional neural networks (CNNs), based on the AlexNet architecture, with a support vector machine (SVM) classifier, applied to multisource images from Sentinel-1 (radar) and Landsat 9 (optical). CNNs are used for edge extraction using Sobel and Laplacian filters, while the SVM ensures precise classification of structural lineaments, roads, faults, and hydrographic networks. The results reveal that Sentinel-1, with its 10-m radar resolution, detects a significantly higher number of lineaments (over 6 million segments) compared to Landsat 9, which provides a more continuous representation of major structures. The comparative analysis highlights the complementarity of the two sensors: Sentinel-1 is more effective in detecting fine and fragmented structures, whereas Landsat 9 promotes spatial coherence of larger lineaments. Validation of the results using Jaccard, Kappa, and coincidence indices, as well as spatial agreement with geological faults and hydrographic networks, confirms the robustness of the approach. This study thus demonstrates the added value of artificial intelligence and multisource integration for improving structural mapping, paving the way for broader applications in geosciences and natural 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

Deep learning-based lineament extraction: a comparative study of sentinel-1, landsat 9 imagery

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Sonia Gannouni,
  • Rihab Riahi,
  • Noamen Rebai

摘要

The automatic detection of geological lineaments from satellite imagery remains a challenge due to the limitations of traditional methods and the complexity of both natural and anthropogenic structures. This study proposes an integrated approach combining convolutional neural networks (CNNs), based on the AlexNet architecture, with a support vector machine (SVM) classifier, applied to multisource images from Sentinel-1 (radar) and Landsat 9 (optical). CNNs are used for edge extraction using Sobel and Laplacian filters, while the SVM ensures precise classification of structural lineaments, roads, faults, and hydrographic networks. The results reveal that Sentinel-1, with its 10-m radar resolution, detects a significantly higher number of lineaments (over 6 million segments) compared to Landsat 9, which provides a more continuous representation of major structures. The comparative analysis highlights the complementarity of the two sensors: Sentinel-1 is more effective in detecting fine and fragmented structures, whereas Landsat 9 promotes spatial coherence of larger lineaments. Validation of the results using Jaccard, Kappa, and coincidence indices, as well as spatial agreement with geological faults and hydrographic networks, confirms the robustness of the approach. This study thus demonstrates the added value of artificial intelligence and multisource integration for improving structural mapping, paving the way for broader applications in geosciences and natural resource management.