English

中国农机化学报

中国农机化学报 ›› 2021, Vol. 42 ›› Issue (12): 155-161.DOI: 10.13733/j.jcam.issn.20955553.2021.12.23

• • 上一篇    下一篇

基于图像特征和随机森林的油菜生物量估算

李海同,陈旭,王刚,关卓怀,江涛,吴崇友   

  1. 农业农村部南京农业机械化研究所,南京市,210014
  • 出版日期:2021-12-15 发布日期:2021-12-15
  • 基金资助:
    中国农业科学院基本科研业务费专项(SR201919);中国农业科学院科技创新工程重大科研任务资助(CAAS—ZDRW202105);国家油菜产业体系专项资助项目(CARS—13—10B);江苏省现代农机装备与技术示范推广项目(NJ2020—11)

Estimation of rapeseed biomass based on image features and random forest

Li Haitong, Chen Xu, Wang Gang, Guan Zhuohuai, Jiang Tao, Wu Chongyou.   

  • Online:2021-12-15 Published:2021-12-15

摘要: 油菜生物量是喂入量和作业质量的主要影响因素,高效、快速地检测油菜生物量是实现油菜收获机自动控制的基础和前提。为研究收获期油菜生物量的影响因素和分布规律,首先利用无人机采集联合收获期油菜的田间可见光图像并实测油菜的生物量信息,提取并构建与油菜生物量有关的32个特征参数,通过相关性分析筛选出与油菜生物量相关性较高的10个显著特征;分别建立基于随机森林(Random forest,RF)、主成分分析(Principal component analysis,PCA)和支持向量机(Support vector machine,SVM)的联合收获期油菜生物量估算模型;利用训练集确定模型参数并优化,利用测试集估算油菜生物量,验证估算模型的性能并比较精度。结果表明:3种模型的评价指标均方根误差(RMSE)、相对误差(RE)和决定系数(R2)分别为0.24 kg/m2、0.04%~22.23%、0.87,0.36 kg/m2、0.92%~21.14%、0.71和0.26 kg/m2、0.28%~34.17%、0.84;对比估算结果可知,基于随机森林的估算模型的RMSE小于PCA和SVM模型,决定系数R2最大且相对误差较小,模型精度和稳定性较优,是估算联合收获期油菜生物量一种较优的方法。基于可见光图像特征和随机森林的油菜生物量估算方法可为油菜联合收割机喂入量自动检测提供方法和参考。

关键词: 无人机, 图像特征, 油菜, 生物量, 随机森林, 估算模型

Abstract:  Rape biomass is one of the most important influencing factors for feed quantity, and operation quality, efficiency, and timely acquisition of feed quantity are of great significance to improve agricultural management. This paper aims to study the influencing factors and distribution law of rape biomass in the harvest period and establishes a prediction model on rape biomass. Firstly, the digital images of rape were taken by Unmanned Aerial Vehicle, and the actual biomass information was measured manually. 32 characteristic parameters related to rape biomass were extracted, and 10 characteristic parameters related to rape biomass were selected according to the significance test results and correlation analysis. With 10 selected features as input set and rape biomass as output, 3 prediction models of rape biomass based on random forest (RF), principal component analysis (PCA), and support vector machine (SVM) were established, respectively. Then, 3 established prediction models were trained using training set data, and parameters for the 3 prediction models were obtained. Finally, the 3 models were used to estimate rape biomass. The root mean square error (RMSE), relative error (RE) and determination coefficient (R2) of three models were 0.24 kg/m2, 0.04%-22.23%, 0.87, 0.36 kg/m2, 0.92%-21.14%, 0.71, and 0.28kg/m2, 0.28%-4.17%, 0.84, respectively. Compared to the estimation results of the 3 models, the RMSE estimation model based on RF was less than the models based on PCA and SVM; R2 was the highest, and its RE was the smallest. Therefore, it is a better method to estimate rape biomass in the combined harvest period. The method of estimating rape biomass based on the UAV digital image proposed in this paper could provide reference and basis for intelligent prediction of feeding quantity in combined rape harvest operation.

Key words:  UAV, image features, rape, biomass, random forest, estimation model

中图分类号: