欢迎访问《草业学报》官方网站,今天是 分享到:

草业学报 ›› 2025, Vol. 34 ›› Issue (1): 41-54.DOI: 10.11686/cyxb2024060

• 研究论文 • 上一篇    下一篇

基于YOLOv5和改进DeeplabV3+的青藏高原植被提取算法

闫储淇(), 黄建强()   

  1. 青海大学计算机技术与应用学院,青海省智能计算与应用实验室,青海 西宁 810016
  • 收稿日期:2024-02-27 修回日期:2024-05-16 出版日期:2025-01-20 发布日期:2024-11-04
  • 通讯作者: 黄建强
  • 作者简介:E-mail: hjqxaly@163.com
    闫储淇(2002-),女,天津人,在读本科。E-mail: 210809010426@qhu.edu.cn
  • 基金资助:
    青海省重点研发计划:地球系统模式公共软件平台在青藏高原气候诊断评估的应用与推广(2023-QY-208)

Vegetation extraction algorithm for the Tibetan Plateau based on YOLOv5 and improved DeeplabV3+

Chu-qi YAN(), Jian-qiang HUANG()   

  1. College of Computer Technology and Application,Qinghai University,Intelligent Computing and Application Laboratory of Qinghai Province,Xining 810016,China
  • Received:2024-02-27 Revised:2024-05-16 Online:2025-01-20 Published:2024-11-04
  • Contact: Jian-qiang HUANG

摘要:

青藏高原的植被覆盖度是生态研究和环境监测的重要指标。传统的植被覆盖度检测方法在地形简单且植被分布集中的区域效果较好,但在复杂地形下由于成本高、调查范围受限、耗时长等问题,导致植被提取精度受限。近年来,计算机视觉和深度学习技术的飞速发展为青藏高原复杂地形下的植被精准提取开辟了新的可能性。本研究提出一种结合YOLOv5和改进DeeplabV3+的双阶段植被提取算法。算法引入基于YOLOv5的植被目标检测模型,以减少背景对第二阶段植被分割任务的干扰;设计新型的DeeplabV3+语义分割模型,以实现精准的植被分割提取。改进的模型引入了轻量级主干网络MobileNetV2、优化了ASPP模块膨胀卷积参数,并集成EMA和CloAttention注意力机制。在青藏高原无人机航拍数据集上的实验结果显示,本算法在交并比(IoU)和像素准确率(PA)上分别达到了90.40%和96.32%,显著超过现有技术,且大幅降低了模型参数。本算法在多种环境条件下均展示了高精度的植被提取能力,可以为青藏高原植被覆盖度的快速、精准测定提供有效的技术支持。

关键词: 青藏高原, 植被提取, 深度学习, YOLOv5, DeeplabV3+

Abstract:

Vegetation coverage on the Qinghai-Tibet Plateau is a crucial metric for ecological studies and environmental monitoring. Traditional methods to detect vegetation coverage are effective in regions with simple terrains and concentrated vegetation. However, in complex terrains, issues such as high costs, restricted survey areas, and extended time intervals reduce the accuracy of the results obtained using such traditional methods. In recent years, rapid advancements in computer vision and deep learning have created new opportunities for precise vegetation extraction in the complex terrains of the Qinghai-Tibet Plateau. Here, we introduce a two-stage vegetation extraction algorithm that integrates YOLOv5 and an improved DeeplabV3+. The algorithm utilizes a vegetation detection model based on YOLOv5 to minimize background interference during the second stage of vegetation segmentation; and a newly designed DeeplabV3+ semantic segmentation model for accurate vegetation segmentation and extraction. The improved model incorporates the lightweight backbone network MobileNetV2, optimizes the dilated convolution parameters of the ASPP module, and integrates EMA and CloAttention mechanisms. The experimental results on the unmanned aerial vehicle dataset of the Qinghai-Tibet Plateau demonstrate that the algorithm attains an intersection over union (IoU) of 90.40% and a pixel accuracy (PA) of 96.32%, significantly outperforming other current technologies and greatly reducing the model’s parameters. Under various environmental conditions, the algorithm exhibits high-precision capabilities for vegetation extraction, offering effective technical support for the rapid and precise measurement of vegetation cover on the Qinghai-Tibet Plateau.

Key words: Tibetan Plateau, vegetation extraction, deep learning, YOLOv5, DeeplabV3+