生物多样性 ›› 2025, Vol. 33 ›› Issue (4): 24236.  DOI: 10.17520/biods.2024236

• 技术与方法 • 上一篇    下一篇

草地植物多样性无人机调查的物种智能识别模型构建

谢淦1,2#,宣晶1,2#,付其迪1#,魏泽1,薛凯1,雒海瑞1,高吉喜3*,李敏1,2*

  

  1. 1中国科学院植物研究所, 北京 100093; 2国家植物园, 北京 100093; 3生态环境部卫星环境应用中心, 北京 100094;
  • 收稿日期:2024-06-14 修回日期:2024-10-14 出版日期:2025-04-20 发布日期:2025-01-03
  • 通讯作者: 李敏

Establishing intelligent identification model for unmanned aerial vehicle surveys in grassland plant diversity

Gan Xie1,2#, Jing Xuan1,2#, Qi-Di Fu1#, Ze Wei1, Kai Xue1, Hai-Rui Luo1, Ji-Xi Gao3*, Min Li1,2*   

  1. 1 Institute of Botany, Chinese Academy of Sciences, Beijing 100093, China

    2 China National Botanical Garden, Beijing 100093, China

    3 Satellite Application Center for Ecology and Environment, Ministry of Ecology and Environment, Beijing 100094, China

  • Received:2024-06-14 Revised:2024-10-14 Online:2025-04-20 Published:2025-01-03
  • Contact: Min Li

摘要: 受全球气候变化和人类活动的影响,植物生存环境、生存状态甚至整个植被生态系统都处于动态变化之中。进行植物的野外定点调查和监测,是科学界和相关管理部门了解和评估区域植物生存状态、推测其发展态势、进而制定相关保护方案等的基础。然而,植物野外调查长期依赖于植物鉴定专家,这意味着极大的人力、物力和财力投入,并且往往难以找到合适的专家。智能识别已被证明能够实现植物的种级精准鉴定,并有极大的潜力能提高植物野外调查的效率、减少专家的工作量。本研究以欧亚草原东段的内蒙古草原为例,通过在呼伦贝尔进行预实验,发现以无人机垂直向下90°拍摄的图片用于构建模型可发现的物种数最多。基于在呼伦贝尔、锡林浩特、鄂尔多斯三地采集的无人机植物图片,利用SSD-MobileNetV2-FPN架构训练无人机植物目标检测模型,对采集图片进行了植物目标检测和提取,利用MobileNetV3架构进行了识图训练,最终搭建了可识别22科54属70种常见草地植物的无人机图像智能识别模型。利用预留的70种10,734幅图片进行评测,该识别模型正确识别了其中的9,513幅,TOP1识别准确率达到88.6%。该模型和无人机图片采集结合,可在80分钟内完成约300个1 m × 1 m样方的调查,极大提升了植物野外调查的效率。本研究为草地植物多样性和生态系统的智能化调查和长期定点监测提供了可用的新工具,并为其他地区和植被类型中开展类似的工作提供可参考的模板。

关键词: 无人机, 人工智能, 物种识别, 草地植物

Abstract

Aims: Under the influence of global climate change and human activities, plant living environment, their living conditions, and even entire vegetation ecosystem are undergoing dynamic changes. Field investigation and monitoring of plants from the foundation for the scientific community and relevant administrative departments to assess the condition of regional plant populations, speculate their development trends, and develop appropriate conservation strategies. However, field plant surveys have long relied on expert identification, requiring significant human, material and financial resources, and finding suitable experts is often challenging. Intelligent identification has been shown to achieve accurate, species-level identification of plants, offering great potential to improve the efficiency of field surveys while reducing the workload of experts.

Methods: In this study, we focused on the eastern part of Eurasia grasslands in Inner Mongolia as an example and employed unmanned aerial vehicle (USVs) along with image recognition technology to develop a grassland plant image recognition model. Pre-experiments conducted in Hulunbeier revealed that images taken by drones at a vertical angle of 90° yielded the highest number of identifiable species for model construction. Based on the drone-captured plant images from Hulunbeier, Xilinhot and Erdos, we used an SSD-MobileNetV2-FPN architecture to train a drone-based plant object detection model. This model detected and extracted plant detection model. Further image recognition training using the MobileNetV3 architecture allowed us to construct an intelligent recognition model capable of identifying 70 common grassland plant species belonging to 54 genera across 22 families.

Results: Using the reserved 10,734 images of 70 species for evaluation, the model correctly identified 9,513 images, achieving a TOP1 recognition accuracy of 88.6%. When combined with UAV-based image acquisition, this model can survey approximately 300 1m*1m quadrats in 80 minutes, significantly improving the efficiency of plant field investigation.

Conclusion: This study provides a new tool for intelligent surveys and long-term site-based monitoring of grassland plant diversity and ecosystems. It reduces the dependence on experts for plant identification, providing a practical tool for biodiversity and ecosystem monitoring at the grassroots level. Additionally, this model offers a template for similar efforts in other regions and vegetation types.

Key words: unmanned aerial vehicle (UAV), artificial intelligence, Species recognition, rangeland plants