Scholarlink

<p>One of the most important challenges for agriculture using aerospace platforms is the identification and zoning of crops. Cultivated area of plantations is used to estimate production, disease detection, water stress assessment among others important monitoring tasks in agriculture. In this work, we proposed a framework to segment the sugarcane crops from high resolution images of a UAV platform. The data acquisition corresponds to the Red, Green, NIR and RedEdge bands of the Parrot Sequoia camera. The first contribution is a binary segmentation strategy as annotation tool, that allows the identification of the sugarcane crop by adaptive thresholding in addition with a supervised labeling in order to generate the training dataset. For segmentation, we adopted three DL CNN models based on U-Net, Attention U-Net and Attention- Residual U-Net architectures, trained with a dataset consisting of 4500 images. The results show that models have similar performance, however, U-Net and Attention U-Net had the highest values of overall accuracy and IoU metrics, even though the latest model required more computation.</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

Semantic Segmentation of Sugarcane Crops in UAV Orthoimages Using U-Net

  • 期刊论文

  • 发表日期：

    2026年4月18日

查看全文
  • Hugo Rene Lárraga Altamirano,
  • Omar Espinosa Guerra,
  • José Tuxpan Vargas,
  • Dulce Carolina Acosta Pintor

摘要

One of the most important challenges for agriculture using aerospace platforms is the identification and zoning of crops. Cultivated area of plantations is used to estimate production, disease detection, water stress assessment among others important monitoring tasks in agriculture. In this work, we proposed a framework to segment the sugarcane crops from high resolution images of a UAV platform. The data acquisition corresponds to the Red, Green, NIR and RedEdge bands of the Parrot Sequoia camera. The first contribution is a binary segmentation strategy as annotation tool, that allows the identification of the sugarcane crop by adaptive thresholding in addition with a supervised labeling in order to generate the training dataset. For segmentation, we adopted three DL CNN models based on U-Net, Attention U-Net and Attention- Residual U-Net architectures, trained with a dataset consisting of 4500 images. The results show that models have similar performance, however, U-Net and Attention U-Net had the highest values of overall accuracy and IoU metrics, even though the latest model required more computation.