欢迎访问《草业学报》官方网站,今天是 分享到:

草业学报 ›› 2024, Vol. 33 ›› Issue (12): 45-58.DOI: 10.11686/cyxb2024045

• 研究论文 • 上一篇    下一篇

无人机多光谱和RGB影像融合的苜蓿产量估测

白宇飞1,2(), 尹航1,2, 杨海波1,2, 冯振华1,2, 李斐1,2()   

  1. 1.内蒙古农业大学草原与资源环境学院,内蒙古自治区土壤质量与养分资源重点实验室,内蒙古 呼和浩特 010018
    2.农业生态安全与绿色发展自治区高等学校重点实验室,内蒙古 呼和浩特 010018
  • 收稿日期:2024-02-01 修回日期:2024-03-07 出版日期:2024-12-20 发布日期:2024-10-09
  • 通讯作者: 李斐
  • 作者简介:E-mail: Lifei@imau.edu.cn
    白宇飞(1997-),女,内蒙古包头人,硕士。E-mail: 2211780885@qq.com
  • 基金资助:
    国家重点研发计划项目(2022YFD1900305-03-01)

Estimation of alfalfa yields on the basis of unmanned aerial vehicle multi-spectral and red-green-blue images

Yu-fei BAI1,2(), Hang YIN1,2, Hai-bo YANG1,2, Zhen-hua FENG1,2, Fei LI1,2()   

  1. 1.College of Grassland,Resources and Environment,Inner Mongolia Agricultural University,Inner Mongolia Key Laboratory of Soil Quality and Nutrient Resources,Hohhot 010018,China
    2.Key Laboratory of Agricultural Ecological Security and Green Development at Universities of Inner Mongolia Autonomous Region,Hohhot 010018,China
  • Received:2024-02-01 Revised:2024-03-07 Online:2024-12-20 Published:2024-10-09
  • Contact: Fei LI

摘要:

产量是苜蓿草场经济产出的重要指标,及时准确量化苜蓿产量有利于提高草场的施肥管理水平和优化种植模式。传统人工草场产量信息的获取依赖于破坏性实地调查,具有滞后性。相比之下,无人机遥感监测技术可以无损地快速获取反演产量模型的有效信息,然而基于单一传感器的图像信息,光谱分辨率和空间分辨率往往不可兼得,从而影响作物长势的全面分析,致使构建的产量模型的估测精度难以得到有效提升。因此,本研究旨在探索无人机多源影像信息融合对收获期苜蓿产量的估测潜力。在苜蓿收获期采集RGB和多光谱(MS)影像,从中提取光谱、纹理以及小波特征,投入偏最小二乘(PLSR)和高斯回归(GPR)两种机器学习算法,评估苜蓿产量估测模型的鲁棒性。结果表明,RGB影像的小波特征相比于颜色指数和纹理特征的估测效果更好,两类特征组合对于苜蓿产量估算精度具有一定的提升,颜色指数、纹理以及小波3类特征结合的GPR苜蓿产量估算模型精度最高,训练集R2=0.76,验证集R2=0.63,分析相对误差(RPD)=1.61。对于MS影像来说,基于纹理特征建立的苜蓿产量估算模型略优于光谱指数特征,组合两类特征构建的GPR苜蓿估产模型精度最高,训练集R2=0.83,验证集R2=0.58,RPD=1.55。当把RGB影像和MS影像特征融合后,苜蓿产量估算模型精度显著提升,融合3类特征参数(多光谱指数、多光谱纹理、RGB小波特征)的GPR模型得到最优的苜蓿产量估算精度,训练集R2=0.83,验证集R2=0.75,RPD=1.98。总体看来,GPR算法表现了最好的估测结果,相比于PLSR模型估测精度提升了13.6%,研究结果为今后标准化人工草地的遥感监测和智能估产提供了参考。

关键词: 苜蓿, 无人机影像, 产量, 植被指数, 纹理特征, 小波特征, 特征融合

Abstract:

Yield is a key component of the economic output of alfalfa (Medicago sativa) pasture. Timely and accurate quantification of alfalfa yield is useful to improve nutrient management and optimize planting patterns. The traditional method to estimate yield in pasture relies on destructive sampling, and there is a certain time lag in obtaining the results. In contrast, unmanned aerial vehicle (UAV)-based monitoring technologies can quickly obtain information to model yield in a non-destructive manner. However, the spectral and spatial resolutions cannot be balanced based on image information from a single sensor, so a comprehensive analysis of crop growth is impossible. It is difficult to effectively improve the accuracy of estimates based on a single UAV image. Therefore, the aim of this study was to explore the potential to combine multi-source image information from UAVs to estimate alfalfa yield during harvesting. In this study, red-green-blue (RGB) and multispectral (MS) images were collected during the alfalfa harvesting period. Based on spectral, texture, and wavelet features extracted from the RGB and MS images, two machine learning algorithms involving partial least squares (PLSR) and Gaussian process regression (GPR) algorithms were used to evaluate the robustness of the alfalfa yield estimation model. The results show that the wavelet features of RGB images were more effective for the comparison of color index and texture features. The combination of the two types of features improved the accuracy of alfalfa yield estimates to some degree. The GPR alfalfa yield estimation model combining three types of features (color index, texture, and wavelet) had high accuracy (training set coefficient of determination R2=0.76, validation set coefficient of determination R2=0.63, and RPD=1.61). For MS images, the model built based on texture features was the most accurate (training set coefficient of determination R2=0.76, validation set coefficient of determination R2=0.63, and the ratio of prediction to deviation RPD=1.61). The alfalfa yield estimation model based on texture features was slightly better than that based on spectral index features, and the GPR alfalfa yield estimation model constructed by combining the two types of features was very accurate (training set coefficient of determination R2=0.83, validation set coefficient of determination R2=0.58, and RPD=1.55). The accuracy of the alfalfa yield estimation model was significantly improved when the RGB image and MS image features were fused. Particularly, the GPR model with three kinds of feature parameters (multi-spectral index, multi-spectral texture, RGB wavelet feature) was the most accurate in estimating alfalfa yield (coefficient of determination R2=0.83 in the training set, coefficient of determination R2=0.75 in the validation set, and RPD=1.98). In conclusion, the GPR algorithm provided the best estimation results, and the estimation accuracy was improved by 13.6% compared with that of the PLSR model. These results provide a reference for remotely monitoring artificial grassland and estimating yield in the future.

Key words: alfalfa, drone image, yield, vegetation index, textural features, wavelet characteristic, feature fusion