摘要
及时、准确的作物估产,对作物管理决策和粮食安全评估具有重要意义。该研究在建立一种耦合连续小波变换(CWT)与机器学习准确预测小麦产量的方法。基于2020年—2021年两年小麦田间试验获取的开花期和灌浆期冠层高光谱数据及产量数据,首先采用CWT提取三种小波特征(WFs),分别为:基于Bortua方法筛选的特征波段(Bortua-WFs)、提取WFs与小麦产量确定系数的前1%(1%R^(2)-WFs)和单一分解尺度下的所有WFs(SS-WFs)。然后采用随机森林(RF)、 K最邻近(KNN)和极端梯度提升(XGBoost)三种机器学习算法构建产量预测模型。最后选取最优的光谱特征,采用相同的方法进行建模并比较。结果表明:(1)三种WFs结合机器学习方法的模型均表现良好,基于Bortua-WFs构建的模型具有更高的精度和稳定性。(2)相比光谱特征模型,Bortua-WFs模型在各生育期的精度均有所提高,开花期的R^(2)精度分别提高了17.5%、 4%和39.6%,灌浆期分别提高了8.4%、 5.6%和16.9%。(3)灌浆期的产量估算模型优于开花期,结合Bortua-WFs和XGBoost的模型表现最佳,R^(2)为0.83, RMSE为0.78 t·ha^(-1)。该研究比较了不同特征和方法相结合的性能,确定了不同方案下的最佳模型精度,为光谱准确预测小麦产量提供技术参考。
Timely and accurate crop yield estimation is crucial for making informed decisions regarding crop management and assessing food security.This study aims to develop a method that combines continuous wavelet transform(CWT)with machine learning to predict wheat yield accurately.This research is based on the spectral data of canopy height and yield data obtained from two-year field trials conducted during wheat growth's flowering and filling stages in 2020—2021.Initially,CWT is employed to extract three wavelet features(WFs),namely Bortua-WFs based on the Bortua method,1%R^(2)-WFs representing WFs along with the top 1%determination coefficient for wheat yield,and SS-WFs encompassing all WFs under a single decomposition scale.Subsequently,three machine learning algorithms Random Forest(RF),K-nearest neighbor(KNN),and extreme gradient Lift(XGBoost)are utilized to construct the yield prediction model.Finally,optimal spectral features are selected using the same methodology for modeling and comparison purposes.The results demonstrate that:(1)all three WFs models combined with machine learning methods perform well,with higher accuracy and stability observed in the model built based on Boruta-WFs.(2)Compared to the spectral characteristic model,improved accuracy was achieved by utilizing Bortua-WFs at each growth stage;specifically,an increase in R^(2) accuracy by 17.5%,4%,and 39.6%during flowering stage,as well as an increase by 8.4%,5.6%,and 16.9%during filling stage respectively were observed across different models.(3)The estimation model at the grouting stage outperformed that at the flowering stage;particularly noteworthy was the performance of XGBoost when combined with Bortua-WFs,which yielded an R^(2) value of 0.83 accompanied by an RMSE value of 0.78 t·ha^(-1).This study compared the performance of different characteristics and methods.It determined the best model accuracy under different schemes,which can provide technical references for the accurate wheat yield prediction by spectral technology.
作者
樊杰杰
邱春霞
樊意广
陈日强
刘杨
边明博
马彦鹏
杨福芹
冯海宽
FAN Jie-jie;QIU Chun-xia;FAN Yi-guang;CHEN Ri-qiang;LIU Yang;BIAN Ming-bo;MA Yan-peng;YANG Fu-qin;FENG Hai-kuan(School of Surveying and Mapping Science and Technology,Xi'an University of Science and Technology,Xi'an 710054,China;Key Laboratory of Quantitative Remote Sensing in Agriculture of Ministry of Agriculture and Rural Affairs,Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;National Engineering and Technology Center for Information Agriculture,Nanjing Agricultural University,Nanjing 210095,China;College of Civil Engineering,Henan University of Engineering,Zhengzhou 451191,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2024年第10期2890-2899,共10页
Spectroscopy and Spectral Analysis
基金
国家重点研发计划项目(2023YFD2000102,2022YFD2001104)
河南省高等学校重点科研项目(24B420004)
河南省科技研发计划联合基金(应用攻关类)项目(222103810024)
河南省科技攻关项目(232102111115)资助。
关键词
连续小波变换
高光谱
机器学习
小麦
产量预测
Continuous wavelet transform
Hyperspectral
Machine learning
Wheat
Yield forecast
