Scholarlink

<p>This study presents a deep network with an encoder-decoder topology for directly extracting a digital terrain model (DTM) from a digital surface model (DSM). Different aspects of the input DSM were extracted at varying levels during the encoding process, while the relevant DTM of the input DSM was extracted during the decoding phase. A multiscale structure was designed to enhance the network’s performance, ensuring the proposed network can effectively remove objects in various settings, particularly in locations with dense foliage and steep slopes. The recommended network displayed excellent performance and exhibited accurate results for DTM extraction during its installation and evaluation in several study regions. The results noted in this study were further compared to the findings of the popular and powerful point cloud filtering methods. Our comparative analysis revealed a significant performance improvement of the proposed network compared to alternative techniques. Specifically, the suggested network achieved an average reduction of 1.74, 0.46, and 0.31 for <i>E</i><sub>RMSE</sub>, <i>E</i><sub>Rel,</sub> and <i>E</i><sub>L</sub> errors, respectively, when compared to LAStools, the second-best performing method.</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

A Multi-scale Convolutional Neural Network for Precise Digital Terrain Model Extraction from Digital Surface Models

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • A’kif Al-Fugara,
  • Ali Nouh Mabdeh,
  • Rami Al shawabkeh,
  • Waleed Al-Khlaifat,
  • Aseel Smerat,
  • Laith Abualigah

摘要

This study presents a deep network with an encoder-decoder topology for directly extracting a digital terrain model (DTM) from a digital surface model (DSM). Different aspects of the input DSM were extracted at varying levels during the encoding process, while the relevant DTM of the input DSM was extracted during the decoding phase. A multiscale structure was designed to enhance the network’s performance, ensuring the proposed network can effectively remove objects in various settings, particularly in locations with dense foliage and steep slopes. The recommended network displayed excellent performance and exhibited accurate results for DTM extraction during its installation and evaluation in several study regions. The results noted in this study were further compared to the findings of the popular and powerful point cloud filtering methods. Our comparative analysis revealed a significant performance improvement of the proposed network compared to alternative techniques. Specifically, the suggested network achieved an average reduction of 1.74, 0.46, and 0.31 for ERMSE, ERel, and EL errors, respectively, when compared to LAStools, the second-best performing method.