等效水厚度梯度的玉米叶片氮素反演模型研究

针对玉米生产中叶片氮素快速、无损检测的实际需求，使用叶级高光谱数据(400～2 500 nm)，依据等效水厚度梯度划分叶片样本，建立了梯度连续的叶片氮素反演模型，初步探索了含水量因素对叶片反射率特性及反演模型精度的影响。首先获取叶级高光谱数据，再根据等效水厚度数值大小对样本进行排序及滑动划分，建立了子集集合。父集除原光谱数据之外还采用了三大类：(1)基线矫正类、(2)散射校正类和(3)平滑处理类光谱变换方法，而子集未使用任何光谱变换方法。建立全波段的PLSR反演模型，对比模型精度，初步定量评价了等效水厚度因素对建模精度的影响。研究结果表明：(1）四组数据中有三组父集反演精度低于最优子集的反演精度，另外一组持平（2018大田低氮：(父)R2_CV=0.48<(子)R2_CV=0.57, (父)RPD_CV=1.38<(子)RPD_CV=1.52；2018大田高氮：(父)R2_CV=0.48<(子)R2_CV=0.7, (父)RPD_CV=1.39<(子)RPD_CV=1.8；2019大田高氮：(父)R2_CV=0.59<(子)R2_CV=0.68, (父)RPD_CV=1.57<(子)RPD_CV=1.77）；(2）四组数据的最优子集反演精度都达到甚至超过了定性模型水平，而父集只有两组；(3）制作反演数据集时在样本筛选问题上需要考虑等效水厚度因素，以避免过于宽泛的样本选择而导致整体反演精度的损失。综上，等效水厚度因素对玉米叶片氮素建模精度存在显著影响，不可忽视。在考虑该因素后，使用叶级高光谱数据对玉米叶片氮素进行快速无损检测的技术方法会更加可信、可行。

Study on Maize Leaf Nitrogen Inversion Model Based on Equivalent Water Thickness Gradient

According to the actual need for rapid and non-destructive testing methods for nitrogen in maize production. Samples were divided according to the equivalent water thickness gradient, and a gradient continuous leaf nitrogen inversion model was established. The influence of water content on leaf reflectance characteristics and the accuracy of the inversion model is preliminarily explored. Firstly, the hyperspectral data of leaf-level are obtained, and then the samples are sorted and sliding divided according to the value of equivalent water thickness, and the subset set is established. In addition to the original spectral data, the parent set also adopts (1) baseline correction；(2) Scattering correction；(3) Smoothing methods, three categories of spectral transformation methods, while subsets do not use any spectral transformation techniques. A full band PLSR inversion model is established, the model accuracy is compared, and the influence of equivalent water thickness on modeling accuracy is preliminarily quantitatively evaluated. The experimental results show that: (1) among the four groups of data, the inversion accuracy of three parent sets is lower than that of the optimal subset, and the other group is the same (2018 field-N: (parent set) R2_CV=0.48<(subset) R2_cv=0.57, RPD_CV=1.38CV=1.52; 2018 field +N: R2_CV=0.48<R2_CV=0.7, RPD_CV=1.39CV=1.8; 2019 field +N: R2_CV=0.59<R2_CV=0.68, RPD_CV=1.57CV=1.77); (2) The inversion accuracy of the optimal subset of all the four groups reaches or even exceeds the level of the qualitative model, while the parent set has only two groups; (3) In the problem of sample selection of inversion data set, the factor of equivalent water thickness needs to be fully considered to avoid the loss of overall inversion accuracy caused by too wide sample selection. In conclusion, the factor of equivalent water thickness significantly impacts the accuracy of nitrogen modeling in maize leaves, which should not be ignored. After this factor is considered, the method of rapid, nondestructive detection of nitrogen in maize leaves using leaf hyperspectral data will be more reliable and feasible.