Scholarlink

<p>Unmanned aerial vehicle (UAV) datasets can derive diverse features, providing crucial support for fine-scale mangrove species classification. However, achieving high classification accuracy remains challenging due to complex feature interactions. This study utilized multi-source UAV data, including multispectral imagery, light detection and ranging (LiDAR) point clouds, and high-resolution RGB images, from the Gaoqiao Mangrove Nature Reserve, Zhanjiang, Guangdong, South China. Three hybrid feature groups were made by integrating shared multispectral features, vegetation indices, and structural features with texture features derived from principal component analysis (PCA), independent component analysis (ICA), or minimum noise fraction (MNF) dimensionality reduction. An improved Extreme Gradient Boosting (XGBoost) algorithm was developed for dominant feature selection, and random forest (RF) and XGBoost models were built for performance evaluation. The optimal results were obtained using PCA features selected by the improved XGBoost algorithm combined with the XGBoost classifier, achieving an overall accuracy of 98.48% with the user accuracy variance of only 0.000 05 among species. These findings indicate that the modified XGBoost algorithm can enhance classification accuracy and robustness, offering technical support for precise mangrove monitoring, protection, and restoration.</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

Improved XGBoost with multi-source UAV data for high-accuracy fine-scale mangrove mapping

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Zhaohui Cheng,
  • Yongze Li,
  • Xiong Sun,
  • Jiajun Yuan,
  • Dazhao Liu,
  • Qinyuan Xiang

摘要

Unmanned aerial vehicle (UAV) datasets can derive diverse features, providing crucial support for fine-scale mangrove species classification. However, achieving high classification accuracy remains challenging due to complex feature interactions. This study utilized multi-source UAV data, including multispectral imagery, light detection and ranging (LiDAR) point clouds, and high-resolution RGB images, from the Gaoqiao Mangrove Nature Reserve, Zhanjiang, Guangdong, South China. Three hybrid feature groups were made by integrating shared multispectral features, vegetation indices, and structural features with texture features derived from principal component analysis (PCA), independent component analysis (ICA), or minimum noise fraction (MNF) dimensionality reduction. An improved Extreme Gradient Boosting (XGBoost) algorithm was developed for dominant feature selection, and random forest (RF) and XGBoost models were built for performance evaluation. The optimal results were obtained using PCA features selected by the improved XGBoost algorithm combined with the XGBoost classifier, achieving an overall accuracy of 98.48% with the user accuracy variance of only 0.000 05 among species. These findings indicate that the modified XGBoost algorithm can enhance classification accuracy and robustness, offering technical support for precise mangrove monitoring, protection, and restoration.