摘要
[目的]作物叶面积指数(LAI)与其长势密切相关,通过研究小麦冠层光谱的不同预处理方法、波段选择方法和模型构建方法的不同组合,找出适用于小麦LAI估算的最佳预测模型,为快速准确监测冬小麦LAI提供参考。[方法]本研究以冬小麦为研究对象,测定其在不同生育时期的LAI与冠层光谱反射率,研究了原始光谱、一阶和二阶导数光谱与LAI之间的相关系数并采用随机蛙跳算法对其进行特征波段的提取,最后基于选取的特征波段,使用偏最小二乘回归(PLSR)和最小二乘支持向量回归(LS-SVR)分别构建LAI预测模型。[结果]结果表明,二阶导数光谱不仅可以改善红边区域波段与LAI之间的相关系数,在725 nm处其相关系数达到0.662,而且提高了特征波段的选择概率,在732 nm处,其选择概率达到0.688。相比采用相关系数和竞争自适应重加权采样(CARS)所建的模型,用随机蛙跳选择的特征波段构建的模型预测精度更高,校正集决定系数达到0.956,验证集决定系数达到0.902,校正集均方根误差降低到0.367,验证集均方根误差降低到0.601。此外,在LAI的预测模型中,LS-SVR的性能优于PLSR。[结论]采用二阶导数预处理结合随机蛙跳特征波长选择算法并使用LS-SVR构建的LAI预测模型性能最佳,可为快速检测LAI提供一种可行的解决方案。
[Objective]The crop leaf area index(LAI)is closely related to its growth vigor.By studying different combinations of different pretreatment methods,band selections methods and model construction methods of wheat canopy spectrum,we expected to find the best prediction model for wheat LAI estimation and provide a reference for rapid and accurate monitoring of wheat LAI.[Methods]In this study,winter wheat was selected as the research object,and its LAI and canopy spectral reflectance at different growth stages were measured.The correlation coefficients between original spectral reflectance,the first and second derivative spectra and LAI were established.The random leapfrog algorithm was used to extract the characteristic wavebands based on three different spectra expressions.The LAI prediction models were constructed with feature wavebands based on partial least squares regression(PLSR)and least squares support vector regression(LS-SVR).[Results]The results indicated that second derivative of spectral reflectance improved the correlation coefficients between waveband reflectance in the red-edge region and LAI,and the correlation coefficient reached 0.662 at 725 nm;the selection probability of the characteristic wavebands was improved up to 0.688 at the waveband of 732 nm.Compared with the model built using correlation coefficients and competitive adaptive re-weighted sampling(CARS),the model constructed with the characteristic bands selected by random leapfrog improved prediction accuracy.The mean determination coefficient reached 0.956 for training set and 0.902 for validation set.The average root mean square error of the calibration set was decreased to 0.367,and 0.601 for validation set compared to those selected by the correlation coefficient and competitive adaptive reweighted sampling.Furthermore,the LSSVR outperformed PLSR in final model of LAI prediction.[Conclusion]The LAI prediction model constructed by using the second derivative preprocessing combined with the random leapfrog characteristic wavelength selection algorithm and LS-SVR produced the best performance.Present results could provide a feasible solution for rapid detection of LAI.
作者
孙晶京
杨武德
冯美臣
肖璐洁
Sun Jingjing;Yang Wude;Feng Meichen;Xiao Lujie(College of Agriculture,Shanxi Agricultural University,Taigu 030801,China;College of Arts and Science,Shanxi Agricultural University,Taigu 030801,China)
出处
《山西农业大学学报（自然科学版）》
CAS
北大核心
2020年第5期120-128,共9页
Journal of Shanxi Agricultural University(Natural Science Edition)
基金
国家自然科学基金(31871571)
国家自然科学基金青年项目(61803245)
山西省科技攻关计划(201903D211002,201603D221037-3)
中国博士后基金(2017M621105)
山西省优秀博士来晋工作奖励资金科研项目(SXYBKY2018040)
山西农业大学科技创新基金(201308)。
