Xinjiang Agricultural Sciences ›› 2024, Vol. 61 ›› Issue (10): 2547-2556.DOI: 10.6048/j.issn.1001-4330.2024.10.022

• Plant Protection · Soil Fertilizer · Water Saving Irrigation · Agricultural Equipment Engineering and Mechanization · Prataculture • Previous Articles     Next Articles

Identification of main plants in desert grassland Seriphidium transiliense based on UAV remote sensing and CNN ResNet 18

QIN Yekangyang(), LI Jiaxin, JIN Guili(), LIU Wenhao, MA Jian, LI Wenxiong, CHEN Mentian   

  1. Xinjiang Key Laboratory of Grassland Resources and Ecology/Key Laboratory of Grassland Resources and Ecology of Western Arid Region, Ministry of Education /College of Grassland Science, Xinjiang Agricultural University, Urumqi 830052, China
  • Received:2024-03-11 Online:2024-10-20 Published:2024-11-07
  • Correspondence author: JIN Guili
  • Supported by:
    Project of National Natural Science Foundation of China(31960360)

基于UAV和CNN ResNet 18参数调节的伊犁绢蒿荒漠草地植物识别性能分析

秦叶康阳(), 李嘉欣, 靳瑰丽(), 刘文昊, 马建, 李文雄, 陈梦甜   

  1. 新疆农业大学草业学院/新疆草地资源与生态重点实验室/西部干旱荒漠区草地资源与生态教育部重点实验室,乌鲁木齐 830052
  • 通讯作者: 靳瑰丽
  • 作者简介:秦叶康阳(2001-),男,江苏沛县人,本科,研究方向为草业科学,(E-mail)2095782457@qq.com
  • 基金资助:
    国家自然科学基金项目(31960360)

Abstract:

【Objective】 Identification of main plants in desert grassland Seriphidium transiliense based on UAV remote sensing and CNN ResNet 18. 【Methods】 In this paper, the Seriphidium transiliense desert grassland, which was concentrated in Xinjiang and at the forefront of degradation, was taken as the research object. The low-altitude UAV remote sensing platform was equipped with a multispectral imager to collect grassland feature information. The ResNet 18 classical model in the deep learning model was selected, too. By setting two groups of training rounds of 40 rounds and 80 rounds, four groups of batch sizes of 8, 16, 32, 64 and five groups of learning rates of 0.01, 0.005, 0.001, 0.000,5, 0.000,1, the model classification performances under different parameter settings were compared and analyzed in order to explore the best parameters for the identification of main species in S. transiliense desert grassland community. 【Results】 The results showed that the overall classification accuracy of S. transiliense community was 83.65% and the classification accuracy of S. transiliense population was 84.21% and the population accuracy of arenarius was 81.15%, when the initial model hyper-parameter was set to 40 training rounds, 8 batch sizes and 0.001 learning rate. By adjusting the model hyper-parameters (the hyper-parameters were set to 40 rounds, the batch size was 32, and the learning rate was 0.0005), the overall classification accuracy of S. transiliense community was 83.73% and the population accuracy of arenarius was 83.78%, and the population accuracy of S. transiliense was 89.18%, which was 0.08% and 4.97% higher than the initial model, respectively. 【Conclusion】 The finding shows that the identification model with high precision, short time and stable performance can be obtained by adjusting the hyper-parameters.

Key words: Seriphidium transiliense; UAV remote sensing; deep learning; ResNet 18; identification

摘要:

【目的】 基于UAV和CNN ResNet 18参数调节的伊犁绢蒿荒漠草地植物识别性能分析。【方法】 以集中分布在新疆且受退化威胁较大的伊犁绢蒿(Seriphidium transiliense)荒漠草地为对象,利用低空无人机遥感平台搭载多光谱成像仪采集该草地地物信息,选择卷积神经网络(Convolutional Neural Networks,CNN)ResNet 18模型,设置40轮和80轮的2组训练轮数,8、16、32、64的4组批量规模,0.01、0.005、0.001、0.000 5、0.000 1的5组学习率3类超参数,对比分析不同参数设置下的模型分类性能,探究适合伊犁绢蒿荒漠草地群落主要物种识别的最佳参数组合。【结果】 初始模型超参数设置为训练轮数40轮、批量规模8、学习率0.001时,伊犁绢蒿群落总体分类精度为83.65%,伊犁绢蒿种群分类精度为84.21%,角果藜(Ceratocarpus arenarius)种群精度为81.15%;通过调节模型超参数(超参数设置为练轮数40轮、批量规模32、学习率0.000 5),伊犁绢蒿群落总体分类精度为83.73%,伊犁绢蒿种群精度为89.18%,角果藜种群精度为83.78%,较初始模型分别提高了0.08%、4.97%和2.63%。【结论】 通过调节超参数可获得精度高、耗时短、性能稳定的伊犁绢蒿荒漠草地植物识别模型。

关键词: 伊犁绢蒿, 无人机遥感, 深度学习, ResNet 18, 识别

CLC Number: