欢迎您访问《智慧农业(中英文)》官方网站! English

Smart Agriculture ?? 2021, Vol. 3 ?? Issue (2): 23-34.doi: 10.12133/j.smartag.2021.3.2.202104-SA003

? ? 上一篇    下一篇

基于无人机图像以及不同机器学习和深度学习模型的小麦倒伏率检测

  

  1. 北达科他州州立大学农业与生物系统工程系,北达科他州法戈市 58102,美国
  • 收稿日期:2021-04-26 修回日期:2021-06-28 出版日期:2021-06-30 发布日期:2021-08-25

Wheat Lodging Ratio Detection Based on UAS Imagery Coupled with Different Machine Learning and Deep Learning Algorithms

Paulo FLORES(), Zhao ZHANG()   

  1. Department of Agricultural and Biosystems Engineering, North Dakota State University, Fargo, ND 58102, USA
  • Received:2021-04-26 Revised:2021-06-28 Online:2021-06-30 Published:2021-08-25
  • corresponding?author: Zhao ZHANG E-mail:paulo.flores@ndsu.edu;zhao.zhang.1@ndsu.edu
  • About author:Paulo FLORES (1979-), male, assistant professor, research interests is remote sensing technologies in agriculture. E-mail: paulo.flores@ndsu.edu.
  • Supported by:
    North Dakota Agricultural Experiment Station Precision Agriculture Graduate Research Assistantship(6064-21660-001-32S);USDA Agricultural Research Service Project(435589)

摘要:

小麦在生长过程中发生倒伏会严重影响其产量,因此实时且准确地对小麦倒伏状况监测有很重要的意义。传统的方法采用手工方式生成数据集,不仅效率低、易出错,而且生成的数据集不准确。针对这一问题,本研究提出了一种基于图像处理的自动数据集生成方法。首先利用无人机在15、46和91 m三个高度采集图像数据;采集完数据后,根据无倒伏、倒伏面积<50%和倒伏面积>50%的标准对每一块地的小麦倒伏情况进行人工评估;采用三种机器学习(支持向量机、随机森林、K近邻)和三种深度学习(ResNet101、GoogLeNet、VGG16)算法对小麦倒伏检测情况进行分类。结果显示,ResNet101的分类结果优于随机森林,并且在91 m高度采集的数据分类精度并不低于在15 m高度采集的数据。本研究证明了针对在91 m高度采集的无人机图像,采用ResNet101对小麦倒伏率检测是一种有效的替代人工检测的方法,其检测精度达到了75%。

关键词: 小麦倒伏率, 机器学习, 深度学习, 数据采集高度, 无人机, ResNet101

Abstract:

Wheat lodging is a negative factor affecting yield production. Obtaining timely and accurate wheat lodging information is critical. Using unmanned aerial systems (UASs) images for wheat lodging detection is a relatively new approach, in which researchers usually apply a manual method for dataset generation consisting of plot images. Considering the manual method being inefficient, inaccurate, and subjective, this study developed a new image processing-based approach for automatically generating individual field plot datasets. Images from wheat field trials at three flight heights (15, 46, and 91 m) were collected and analyzed using machine learning (support vector machine, random forest, and K nearest neighbors) and deep learning (ResNet101, GoogLeNet, and VGG16) algorithms to test their performances on detecting levels of wheat lodging percentages: non- (0%), light (<50%), and severe (>50%) lodging. The results indicated that the images collected at 91 m (2.5 cm/pixel) flight height could yield a similar, even slightly higher, detection accuracy over the images collected at 46 m (1.2 cm/pixel) and 15 m (0.4 cm/pixel) UAS mission heights. Comparison of random forest and ResNet101 model results showed that ResNet101 resulted in more satisfactory performance (75% accuracy) with higher accuracy over random forest (71% accuracy). Thus, ResNet101 is a suitable model for wheat lodging ratio detection. This study recommends that UASs images collected at the height of about 91 m (2.5 cm/pixel resolution) coupled with ResNet101 model is a useful and efficient approach for wheat lodging ratio detection.

Key words: wheat lodging ratio, machine learning, deep learning, mission height, UAS, ResNet101

中图分类号: 

撸撸色在线看观看免费 <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <蜘蛛词>| <文本链> <文本链> <文本链> <文本链> <文本链> <文本链>