基于赤池信息准则的冬小麦植株氮含量高光谱估算

Hyperspectral estimation of plant nitrogen content based on Akaike's information criterion

为了快速、准确地测定冬小麦植株氮含量,利用2014?2015年的冬小麦冠层反射光谱数据构建了16种氮素或叶绿素敏感光谱指数,基于变量投影重要性(variable importance projection,VIP)-偏最小二乘(partial least squares,PLS)-赤池信息准则(Akaike’s information criterion,AIC)整合模型构建了不同生育期植株氮含量最佳回归模型,并用2012?2013年挑旗期数据对模型进行了验证。结果表明:在AIC下,拔节期以4个植被指数为自变量的模型最优;挑旗期以5个植被指数为自变量的模型最优;开花期以4个植被指数为自变量的模型最优;灌浆期以6个植被指数为自变量的模型最优。4个生育期建模的决定系数(R2)和均方根误差(RMSE)分别为0.71、0.86、0.75、0.46和0.23%、0.13%、0.12%、0.15%,以挑旗期决定系数为最大。挑旗期验证集的R2和RMSE分别为0.81和0.41%,预测模型和验证模型均具有较高的估算精度和可靠性,研究结果为选择小麦合适的生育期估算小麦植株氮营养状况提供参考。

In order to measure plant nitrogen content（PNC） rapidly and accurately in different growth stages, the optimal regression model for PNC was constructed based on variable importance projection – partial least squares – Akaike＇s information criteria（VIP-PLS-AIC） and corresponding PNC data. In this research, 16 spectral indices sensitive to nitrogen and chlorophyll were constructed by using of winter wheat canopy reflectance obtained in National Precision Agriculture Experimental Base from 2014 to 2015. The model was verified by using of data at flag leaf stage from 2012 to 2013. Results showed that in jointing stage the related degree order between VIP evaluation sixteen vegetation index and winter wheat PNC can be drawn as follows： PPR 〉RedWidth〉 SRPI〉 NPCI〉 NPQI 〉SIPI 〉RedArea 〉MCARI/MTVI2〉 TCARI 〉PSNDc 〉MCARI 〉DCNI 〉REPGAUSS〉 REP〉 PRI〉 SR（533,565）. In booting stage the order between VIP and PNC can be drawn as follows： PPR 〉SRPI 〉NPCI〉 NPQI〉 MCARI/MTVI2〉 SR（533,565）〉 PRI〉 SIPI〉REPGUSS〉REP〉RedArea〉PSNDc〉Red Width〉DCNI〉MCARI〉TCARI. In anthesis stage the order between VIP and PNC can be described as PPR 〉NPQI〉 MCARI 〉MCARI/MTVI2〉 TCARI 〉DCNI 〉REPGAUSS 〉REP〉 SR（533,565） 〉SRPI 〉NPCI〉 PSNDc〉 RedWidth 〉PRI 〉RedArea〉 SIPI. In filling stage, the order between VIP and PNC can be described as TCARI〉 MCARI〉 NPQI 〉DCNI〉 SIPI〉 MCARI/MTVI2〉 PPR 〉RedArea 〉REPGAUSS 〉REP 〉PSNDc〉 RedWidth〉 SR（533,565） 〉PRI 〉SRPI〉 NPCI. The PNC model of winter wheat based on AIC at jointing stage using four vegetation indices as independent variables was the optimal. At flag leaf stage, flowering stage and filling stage they were five, four and six kinds, respectively. The determined coefficients（R^2） and root mean square error（RMSE） during four growth stages were 0.71, 0.86, 0.75, 0.46 and 0.23%、0.13%、0.12%、0.15%, respectively. At booting stage the independent variables respectively were VPPR, VSRPI, VNPCI, VNPQI and VMCARI/MTVI2. The booting stage in 2012 to 2013 data was used to validate and the booting stage was the optimal stage for estimating winter wheat PNC using hyperspectral data. The results showed R^2 and RMSE of validation set at booting stage were 0.81 and 0.41%. Besides, both prediction model and verification model had higher accuracy and reliability. The estimation result of winter wheat PNC based on coupling model VIP-PLS-AIC was ideal and provided an effective method for predicting winter wheat PNC by remote sensing. The overall results showed that the PNC of winter wheat can be reliably monitored with the canopy spectral methods established in the study.