机器学习结合高光谱植被指数与SPAD值估算冬小麦氮含量

冬小麦叶片氮含量与叶片光合作用和营养状况密切相关，直接影响植株生长发育，而茎秆中的氮含量与茎秆中纤维素、半纤维素和木质素的比例和含量密切相关，直接影响茎秆质量及植株的抗倒伏能力。然而，有关对冬小麦茎秆氮含量估算研究较为有限，限制了从氮含量角度判断茎秆质量及对倒伏的预测能力。为精准估算冬小麦不同器官（叶片、茎秆）氮含量，该研究通过2年田间试验，获取冬小麦4个关键生育期（拔节期、抽穗期、开花期、灌浆期）和3种施氮水平条件下（N1、N2和N3）的冠层光谱反射率、叶片、茎秆氮含量及叶片SPAD (soil and plant analyzer development， SPAD)值。分析了不同生育期和施氮水平条件下高光谱植被指数对叶片和茎秆氮含量的敏感性，并结合5种常用的机器学习算法：随机森林回归（random forest regression，RFR）、支持向量回归（support vector regression，SVR）、偏最小二乘回归（partial least squares regression，PLSR）、高斯过程回归（gaussian process regression，GPR）、深度神经网络回归（deep neural networks，DNN）构建冬小麦叶片和茎秆氮含量估算模型。结果表明：高光谱植被指数对叶片和茎秆氮含量的敏感性受到生育期和施氮水平的影响。在灌浆期，最佳植被指数双峰冠层植被指数 DCNI（double-peak canopy nitrogen index）对叶片氮含量的敏感性最高，R²为0.866。对茎秆氮含量，在抽穗期的敏感性最高，最佳植被指数归一化叶绿素比值指数 NPQI（normalized phaeophytinization index）与氮含量相关系数R²＝0.677。施氮水平的提升增加了光谱植被指数对茎秆氮含量的敏感性。结合SPAD值的机器学习算法提升了氮含量的估算精度，对叶片氮含量，在不同生育期和施氮水平条件下估算精度提升了1%～7%，其中在全生育期的归一化均方根误差NRMSE从0.254提升到0.214，抽穗期的NRMSE提升最大，从0.201提升到0.128。对茎秆氮含量，全生育期的NRMSE从0.443提升到0.400，抽穗期的NRMSE提升最大，从0.323提升到0.268。在全生育期，结合SPAD值的DNN模型对叶片（R²=0.782、NRMSE=0.214）和茎秆（R²=0.802、NRMSE=0.400）氮含量的估算精度最佳。研究说明，SPAD值与光谱植被指数结合有利于提升冬小麦不同生育期和施氮水平条件下叶片和茎秆氮含量的估算精度。

Estimating winter wheat nitrogen content using SPAD and hyperspectral vegetation indices with machine learning

Leaf nitrogen content is closely related to leaf photosynthesis and the nutritional status of winter wheat plants, which directly affects the plant growth and development. While the stem nitrogen content is closely related to the proportion and content of cellulose, hemicellulose and lignin in stems, which directly affects stem quality and plant lodging resistance. However, it is still lacking on the direct estimation of stem N content in winter wheat. It is very necessary to evaluate the stalk quality and predict lodging from the perspective of stem nitrogen content. In this study, a 2-year field experiment was conducted to accurately estimate the nitrogen content in different plant organs (leaves and stems) of winter wheat. Winter wheat canopy spectral reflectance, leaf and stem nitrogen content, and leaf SPAD (Soil and Plant Analyzer Development) values were obtained at four growth stages (jointing, heading, anthesis and filling) and three nitrogen application levels (N1, N2 and N3). A systematic analysis was made to determine the sensitivity of hyperspectral vegetation indices to leaf and stem nitrogen contents at different growth stages and nitrogen application levels. Five commonly-used machine learning algorithms were used to estimate the leaf and stem nitrogen contents of winter wheat, including random forest regression (RFR), support vector regression (SVR), partial least squares regression (SVR), partial least squares regression (PLSR), Gaussian process regression (GPR) and deep neural networks (DNN). The hyperspectral vegetation indices only or combined with SPAD were used as the inputs to construct the nitrogen estimation models for leaves and stems. The results showed that the sensitivity of hyperspectral vegetation indices to the nitrogen content in leaves and stems was influenced by the growth stage and nitrogen application level. In the filling stage, the best vegetation index DCNI (Double-peak canopy nitrogen index) shared the highest sensitivity to leaf nitrogen content, where the determination coefficient R² was 0.866. The sensitivity to stem nitrogen content was the highest at the heading stage, and the R² between the best vegetation index NPQI (Normalized phaeophytinization index) and nitrogen content was 0.677. The sensitivity of the spectral vegetation index to the stem nitrogen content increased with the increasing nitrogen application level. The machine learning combined with the SPAD value and vegetation indices was improved the estimation accuracy of the nitrogen content, compared with only the vegetation index. In leaf nitrogen content, the estimation accuracy increased by 1%-7% under different growth stages and nitrogen application levels. The normalized root mean square error (NRMSE) increased from 0.254 to 0.214 during the whole growth stage. In a single growth period, the NRMSE increased from 0.201 to 0.128 at the heading stage, indicating the most increase. In the stem nitrogen content, the NRMSE increased from 0.443 to 0.400 during the whole growth stage. There was the most increase during the heading stage with the values ranging from 0.323 to 0.268. In the whole growth period, the DNN model combined with the SPAD value was achieved the best accuracy to estimate the nitrogen content of leaves (R²=0.782 and NRMSE=0.214) and stems (R²=0.802 and NRMSE=0.400). The combination of the SPAD value and spectral vegetation index can be expected to improve the accuracy of the nitrogen content in the leaves and stems of winter wheat at different growth stages and nitrogen application levels.