Scholarlink

<p>Mapping young forests is helpful for future forest management, including monitoring forest growth, assessing regeneration, and supporting effective silvicultural management strategies. Convolutional neural networks (CNNs) enable rapid and efficient tree mapping when combined with remote sensing imagery. In this study, we compared the classification performance of Mask Region-based CNN and Faster Region-based CNN using various UAV image resolutions for a young <i>Casuarina equisetifolia</i> L. forest located in Pingtan Comprehensive Pilot Zone, Fujian, China. The red, green, and blue band imagery obtained from a DJI Matrice 300 Real-Time Kinematic and a ZENMUSE P1 camera was resampled to images with 1.00, 2.00, 3.00, and 4.00&#xa0;cm resolutions. Tree crowns were delineated from the original red, green, and blue band imagery, ensuring each polygon corresponded to a single tree on the ground. A total of 2273 tree crown polygons were used for model training, and the remaining 2948 tree crowns were used as the test set for model accuracy evaluation. The results showed that the highest accuracy was achieved using Mask Region-based CNN with an image resolution of 2.00&#xa0;cm (F1 score = 95.18%, IoU = 74.03%), followed by the Faster Region-based CNN with an image resolution of 1.00&#xa0;cm (F1 score = 84.62%, IoU = 70.95%). The optimal image resolution differs for each model to achieve the best performance. The study highlights that high resolution images may not be optimal for individual tree detection when applying CNN models. It also provides valuable insights for selecting appropriate CNN models and image resolutions for tree detection.</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

Comparative performance of convolutional neural networks for detecting and mapping a young Casuarina equisetifolia L. forest from unmanned aerial vehicle (UAV) imagery

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Zhenbang Hao,
  • Shilong Yao,
  • Christopher J. Post,
  • Elena A. Mikhailova,
  • Lili Lin

摘要

Mapping young forests is helpful for future forest management, including monitoring forest growth, assessing regeneration, and supporting effective silvicultural management strategies. Convolutional neural networks (CNNs) enable rapid and efficient tree mapping when combined with remote sensing imagery. In this study, we compared the classification performance of Mask Region-based CNN and Faster Region-based CNN using various UAV image resolutions for a young Casuarina equisetifolia L. forest located in Pingtan Comprehensive Pilot Zone, Fujian, China. The red, green, and blue band imagery obtained from a DJI Matrice 300 Real-Time Kinematic and a ZENMUSE P1 camera was resampled to images with 1.00, 2.00, 3.00, and 4.00 cm resolutions. Tree crowns were delineated from the original red, green, and blue band imagery, ensuring each polygon corresponded to a single tree on the ground. A total of 2273 tree crown polygons were used for model training, and the remaining 2948 tree crowns were used as the test set for model accuracy evaluation. The results showed that the highest accuracy was achieved using Mask Region-based CNN with an image resolution of 2.00 cm (F1 score = 95.18%, IoU = 74.03%), followed by the Faster Region-based CNN with an image resolution of 1.00 cm (F1 score = 84.62%, IoU = 70.95%). The optimal image resolution differs for each model to achieve the best performance. The study highlights that high resolution images may not be optimal for individual tree detection when applying CNN models. It also provides valuable insights for selecting appropriate CNN models and image resolutions for tree detection.