基于高光谱和数据挖掘的油菜植株含水率定量监测模型  

Quantitative monitoring models of plant water content in rapeseed based on hyperspectrum and related data mining

作　　者：潘月 曹宏鑫[2] 齐家国 吴菲 韩旭杰 丁昊迪 葛道阔[2] 张玲玲 张伟欣[2] 张文宇[2] PAN Yue;CAO Hong-xin;QI Jia-guo;WU Fei;HAN Xu-jie;DING Hao-di;GE Dao-kuo;ZHANG Ling-ling;ZHANG Wei-xin;ZHANG Wen-yu(College of Agriculture/Asia Hub on Agriculture,Nanjing Agricultural University,Nanjing 210095,China;Institute of Agricultural Information,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China;College of Agriculture,Yangzhou University,Yangzhou 225009,China)

机构地区：[1]南京农业大学农学院/亚洲农业研究中心,江苏南京210095 [2]江苏省农业科学院农业信息研究所,江苏南京210014 [3]扬州大学农学院,江苏扬州225009

出　　处：《江苏农业学报》2022年第6期1550-1558,共9页Jiangsu Journal of Agricultural Sciences

基　　金：国家自然科学基金项目(31471415、31871522);江苏省农业科技自主创新资金项目[CX(19)2040-1]。

摘　　要：为了构建监测效果更好、更具普适性的油菜植株含水率(Plant water content,PWC)定量监测模型,以油菜品种浙杂903、宁油22和宁杂1818为试验材料,设置2个施肥水平和3个水分处理,基于2019-2020年和2020-2021年生长季田间试验资料,在PWC的高光谱响应敏感波段范围采用逐步回归(Stepwise regression, SR)分析、连续投影算法(Successive projection algorithm, SPA)、竞争自适应加权算法(Competitive adaptive reweighted sampling, CARS)以及减量精细采样法(Reduced precise sampling method, RPSM)深度挖掘高光谱数据,通过筛选最优波段组合与光谱指数,基于线性回归(Linear regression, LR)、BP神经网络(Back-propagation neural network, BPNN)和支持向量机回归(Support vector regression, SVR)方法构建并比较油菜植株含水率监测模型。结果表明,针对油菜PWC监测,SR分析筛选的最优波段组合为730 nm、986 nm和1 071 nm, SPA法分析筛选的最优波段组合为686 nm、695 nm、707 nm、746 nm、964 nm、1 065 nm和1 069 nm, CARS法分析筛选的最优波段组合为694 nm、695 nm、696 nm、863 nm、864 nm、893 nm、973 nm、986 nm、1 050 nm和1 071 nm。RPSM筛选的最优光谱指数是归一化差值光谱指数(R981,R894)和比值光谱指数(R981,R894),其利用的波段均位于近红外波段。前述3个方法筛选的波段变量更多,蕴含的信息更全面,估测精度普遍优于光谱指数。建模分析结果表明,SPA-LR模型、SPA-BP模型、SPA-SVR模型均能实现油菜PWC的精确监测,经检验,其估测值和实测值的R2分别为0.693、0.940、0.841,均方根误差(RMSE)分别为1.623%、1.836%和1.227%。结果证明高光谱数据具备深度挖掘价值,运用全波段光谱分析方法能够在降维的同时保留有效信息,利用筛选出的波段组合构建线性或非线性模型,均能实现大田条件下全生育期油菜植株含水率的定量监测。To construct a quantitative monitoring model for plant water content(PWC) of rapeseed with relative better monitoring effect and more universality, rapeseed cultivars Zheza 903, Ningyou 22, and Ningza 1818 were used as the experimental materials in this study, two fertilization levels and three water treatments were set. Based on field test data in growing seasons of 2019-2020 and 2020-2021, the hyperspectral data were deeply minined by stepwise regression(SR) analysis, successive projection algorithm(SPA), competitive adaptive reweighted sampling(CARS) and reduced precise sampling method(RPSM), within the sensitive band range of hyperspectral response of PWC. By screening the optimal band combination and spectral index, the monitoring models of the rapeseed PWC were constructed and compared based on linear regression(LR), back-propagation neural network(BPNN), and support vector machine regression(SVR). The results showed that, for the monitoring of rapeseed PWC, the optimal bands combination by SR analysis was 730 nm, 986 nm and 1 071 nm, the optimal bands combination by SPA method was 686 nm, 695 nm, 707 nm, 746 nm, 964 nm, 1 065 nm and 1 069 nm, and the optimal bands combination by CARS method was 694 nm, 695nm, 696 nm, 863 nm, 864 nm, 893 nm, 973 nm, 986 nm, 1 050 nm and 1 071 nm. The optimal spectral indices screened by RPSM were reduced precise sampling method(NDSI)(R981, R894) and ratio spectral index(RSI)(R981, R894), all the utilized bands were located in the near-infrared band region. The above three methods screened relatively more band variables, contained relatively more comprehensive information, and the estimation accuracies were generally better than spectral indices. Analysis results of modeling showed that, the SPA-LR model, SPA-BP model and SPA-SVR model could realize accurate monitoring of rapeseed PWC. Through testing, R~2 of estimated value and measured value were 0.693, 0.940 and 0.841, respectively, and root mean square error(RMSE) were 1.623%, 1.836% and 1.227%, respectively. This study

关 键 词：高光谱 油菜 连续投影算法 竞争自适应加权算法 BP神经网络 

分 类 号：S126[农业科学—农业基础科学]