基于CARS特征波段筛选的冬小麦植株氮浓度监测

Nitrogen concentration monitoring in winter wheat plants based on CARS characteristic band screening

关键词:  冬小麦  植株氮含量  CARS  冠层氮指数  高光谱

Key words:  winter wheat  plant nitrogen content  CARS  canopy nitrogen index  hyperspectral

摘要:

【目的】 氮素在作物生长发育、产量及品质形成中不可或缺的营养元素。高效、无损、精准地获取作物氮素盈亏状况，能够监测作物长势，提高氮肥施用水平和利用效率，降低施肥过量导致的农田面源污染。【方法】 文章对2020—2022年3年高光谱数据进行SG平滑、一阶导数预处理。将相关性分析（Correlation analysis，CA）与竞争性自适应重加权采样法（Competitive Adaptive Reweighted Sampling，CARS）相结合（CA-CARS），研究光谱一阶导数与植株氮浓度（Plant nitrogen concentration，PNC）的关系，明确拔节期不同氮素处理下的敏感性波段。最终筛选出最敏感波段构建植被指数，基于此建立冬小麦植株氮浓度一元线性监测模型。以2020年、2022年数据为训练集建模、2021年数据为验证集进行模型精度验证。【结果】 （1）综合3年拔节期不同氮水平下，冬小麦PNC高度敏感波段区位主要有：蓝绿波段（495 nm~503 nm）、红边范围（736 nm~750 nm）及近红外范围（751 nm~753 nm、751 nm~753 nm、761 nm~765 nm、773 nm~779 nm、922 nm、937 nm~938 nm、1 016 nm~1 032 nm、1 083 nm~1 088 nm、1 127 nm、1 142 nm~1 145 nm、1 292 nm~1 300 nm）。（2）CARS筛选出6个特征波段为459 nm、682 nm、721 nm、746 nm、1 049 nm、1 175 nm。（3）利用特征波段组建15个冠层比值氮指数（Canopy Ratio Nitrogen Index，CRNI），CRNI10的模型精度最高、均方根误差最小。其训练集验证集决定系数、均方根误差分别为R²=0.785、R²=0.679、RMSE=0.254和RMSE=0.332。说明该文构建的CRNI在PNC监测上更具泛化性。【结论】 通过CA-CARS结合的方式筛选出的特征参数所构建的PNC反演模型，能有效提升PNC监测模型的精度、迁移性及稳定性。

Abstract:

Purpose Nitrogen is an essential nutrient element for crop growth，development，yield，and quality. Efficient，non-destructive and accurate assessment of crop nitrogen status enables monitoring of crop growth，enhances nitrogen fertilizer application levels and efficiency，and reduces agricultural non-point source pollution resulting from excessive fertilizer application.Method This study conducted SG smoothing and first-order derivative preprocessing on three years of hyperspectral data（2020—2022）. Correlation analysis（CA）and competitive adaptive reweighted sampling（CARS）（CA-CARS）were applied to investigate the relationship between the first-order derivative of the spectrum and plant nitrogen concentration（PNC），identifying the sensitive spectral bands under different nitrogen treatments at the jointing stage. The most sensitive bands were then selected to construct a vegetation index，upon which a univariate linear monitoring model for winter wheat PNC was developed. The data from 2020 and 2022 were used as a training set for modeling，while the data from 2021 served as the validation set for model accuracy assessment.Result （1）The study showed that under different nitrogen levels in the integrated three-year nodulation period，the highly sensitive band locations of winter wheat PNC were mainly the blue-green band（495 nm-503 nm），the red-edge range（736 nm-750 nm），and the near-infrared range（751 nm-753 nm，751 nm-753 nm，761 nm-765 nm，773 nm-779 nm，922 nm，937 nm-938 nm，1 016 nm-1 032 nm，1 083 nm-1 088 nm，1 127 nm，1 142 nm-1 145 nm，1 292 nm-1 300 nm）.（2）CARS identified 6 characteristic bands，specifically 459 nm，682 nm，721 nm，746 nm，1 049 nm，and 1 175 nm. （3）A total of 15 canopy ratio nitrogen indices（CRNI）were constructed using these characteristic bands，with CRNI10 exhibiting the highest model accuracy and the lowest root mean square error（RMSE）. The coefficient of determination（R2）and RMSE for the training and validation sets were R²=0.785，R²=0.679，RMSE=0.254 and RMSE=0.332，respectively. These results indicated that the CRNI model developed in this study was more generalizable in PNC monitoring.Conclusion The PNC inversion model constructed by the feature parameters screened by the combination of CA-CARS can effectively improve the accuracy，migration and stability of the PNC monitoring model.