光谱学与光谱分析, 2020, 40 (7): 2259, 网络出版: 2020-12-05   

基于模型集群的马铃薯叶绿素检测光谱变量筛选讨论

Discussion on Spectral Variables Selection of Potato Chlorophyll Using Model Population Analysis
作者单位
1 中国农业大学现代精细农业系统集成研究教育部重点实验室, 北京 100083
2 Center for Precision & Automated Agricultural System, Washington State University, Pullman WA 99350, USA
摘要
为了探究马铃薯作物叶绿素吸收特征, 充分解析光谱特征波长变量, 建立高精度叶绿素含量检测模型。 在马铃薯发棵期(M1)、 块茎形成期(M2)、 块茎膨大期(M3)和淀粉积累期(M4)4 个关键生长期, 利用ASD便携式光谱仪采集80个样本区的314组作物冠层反射率数据, 并同步采集叶片测定叶绿素含量。 在光谱数据预处理之后, 分析了马铃薯不同生长期的光谱反射率变化特征。 利用基于模型集群思想的蒙特卡洛无信息变量消除(MC-UVE)、 随机蛙跳(RF)、 竞争自适应重加权采样(CARS)三种算法筛选叶绿素特征波长, 建立叶绿素含量检测PLS模型。 对4个生长期的314个样本, 采用SPXY算法分别按照3∶1的比例划分, 得到建模集240个样本、 验证集74个样本。 利用MC-UVE, RF, CARS三种算法筛选叶绿素特征波长, 讨论迭代次数(N)和特征变量个数(LV)对MC-UVE和RF算法、 迭代次数(N)对CARS算法筛选特征波长结果的影响, 对迭代次数设置6个梯度, 分别为N=50, 100, 500, 1 000, 5 000和10 000; 对特征变量数设置4个梯度, 分别为LV=15, 20, 25和30。 以PLSR模型的验证集结果为评价指标, 分析迭代次数(N)和特征变量数(LV)的最优参数组合。 最后基于MC-UVE, RF和CARS算法筛选得到的最佳特征波长建立叶绿素检测PLSR模型, 分别记为MC-UVE-PLSR, RF-PLSR, CARS-PLSR。 结果表明, CARS, RF和MC-UVE三种算法的迭代次数(N)、 特征变量数(LV)参数最佳组合分别为: (1)MC-UVE: 迭代次数N=50 特征变量数LV=30; (2)RF: 迭代次数N=500、 特征变量数LV=30; (3)CARS: 迭代次数N=100。 对比在最佳特征波长建立的MC-UVE-PLSR, RF-PLSR, CARS-PLSR叶绿素含量检测, 发现RF-PLSRRR模型的性能最优, R2v为0.786, RMSEV为3.415 mg·L-1; MC-UVE-PLS模型性能次之, R2v为0.696, RMSEV为4.072 mg·L-1; CARS-PLS模型的性能最差, R2v为0.689, RMSEV为4.183 mg·L-1。 以上结果说明: 在筛选马铃薯叶绿素特征波长方面RF算法优于MC-UVE和CARS, 得到的特征波长能够较全面地反映与马铃薯叶绿素相关的物质信息。
Abstract
The paper was aimed to explore the chlorophyll spectral absorption characteristics of potato crops, fully analyze the spectral characteristic wavelength variables, and establish a high--precision chlorophyll content detection model. The 314 reflectance samples were collected using an ASD portable spectrometer at the seedling stage (M1), tuber formation stage (M2), tuber expansion stage (M3) and starch accumulation stage (M4). The chlorophyll content was determined by the simultaneous collection of leaves. After spectral data pre--treatment, the spectral reflectance changes of different growth stages of potato were analyzed. The algorithms based on model population analysis were used to select chlorophyll characteristiccharacteristic chlorophyll wavelengths, including Monte Carlo uninformative variables elimination (MC--UVE), random frog (RF) and competitive adaptive reweighted sampling (CARS) algorithm. The partial least square regression (PLSR) was used to establish the chlorophyll content detection model. The sample set was divided by a ratio of 3∶1 in each growth stage using the sample set partitioning based on joint X-Y distance algorithm (SPXY) with the 240 calibration samples and 74 validation samples. The different algorithms (MC-UVE, RF, CARS) were used to select chlorophyll characteristic wavelengths. The influence of the number of iteration (N) and the number of the latent variables (LV) on the results of characteristic wavelength selection of MC-UVE and RF algorithms were discussed, and the influences of N on that of CARS algorithm were discussed. Six gradients were set for the number of iterations (N), which were N=50, 100, 500, 1 000, 5 000 and 10 000, respectively. Four gradients were set for the number of latent variables (LV), which were LV=15, 20, 25 and 30 respectively. Taking the validation set result of PLS model as the evaluation index, the optimal parameter combination of N and LV was analyzed. Based on the optimal characteristic wavelengths selected by the three algorithms, the chlorophyll detection PLSR models were established and denoted as RF-PLSR, MC-UVE-PLSR, and CARS-PLSR, respectively. The research results showed that the chlorophyll characteristic wavelengths selection results were optimal when N=50 and LV=30 of MC-UVE, N=500 and LV=30 of RF, N=100 of CARS. By comparing the RF-PLSR, MC-UVE-PLSR, and CARS-PLSR models, it was indicated that the performance of the RF-PLSR model was best, the determination coefficient of validation (R2v) was 0.786, the root means square error of validation (RMSEV) was 3.415 mg·L-1; MC-UVE-PLSR was second, the R2v was 0.696, the RMSEV was 4.072; and the CARS-PLSR was the worst, the R2v was 0.689, the RMSEV was 4.183. Above results showed that the RF algorithm was superior to MC-UVE and CARS in selecting the characteristic chlorophyll wavelength of potato.

刘宁, 邢子正, 乔浪, 李民赞, 孙红, Qin Zhang. 基于模型集群的马铃薯叶绿素检测光谱变量筛选讨论[J]. 光谱学与光谱分析, 2020, 40(7): 2259. LIU Ning, XING Zi-zheng, QIAO Lang, LI Min-zan, SUN Hong, Qin Zhang. Discussion on Spectral Variables Selection of Potato Chlorophyll Using Model Population Analysis[J]. Spectroscopy and Spectral Analysis, 2020, 40(7): 2259.

本文已被 1 篇论文引用
被引统计数据来源于中国光学期刊网
引用该论文: TXT   |   EndNote

相关论文

加载中...

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!