Abstract:

In order to construct a hyperspectral monitoring model of nitrogen content in the canopy of multiple varieties of millet， the hyperspectral reflectance and leaf nitrogen content of millet in the whole growth stage of millet were obtained by setting up different nitrogen levels and field experiments of multiple varieties of millet. The data of hyperspectral reflectance were preprocessed by convolution smoothing and first derivative transformation. The correlation between hyperspectral data and leaf nitrogen content of millet was analyzed. Nitrogen sensitive bands， vegetation index and hyperspectral characteristic parameters of millet at different and whole growth stages were screened by successive projections algorithm （SPA） and correlation analysis between spectral data and nitrogen content. The partial least square regression （PLSR） estimation model of nitrogen content in millet canopy was established by combining the 3 combinations. The results showed that the optimal estimation models of different growth periods were different. At jointing stage，the model based on sensitive band，vegetation index and hyperspectral characteristic parameters had the highest accuracy. At heading stage， the model based on sensitive band， vegetation index model had the highest accuracy. At pustulation stage， the model based on vegetation index and hyperspectral characteristic parameter model had the highest accuracy. At mature stage， the model based on sensitive band， and vegetation index model had the highest accuracy. In the whole growth period， the model based on sensitive bands and vegetation index had the highest accuracy. The multi-input level synthesis model could make full use of the spectral information to effectively improve the prediction accuracy and stability of the model， and the model based on sensitive band and vegetation index performs had best effect， which the R² of prediction set was more than 0.82， root-mean-square error （RMSE） was less than 0.119， and relative predicted deviation （RPD） was greater than 2.1. Above results provided theoretical basis and technical support for hyperspectral remote sensing to diagnose nitrogen surplus and deficiency and fertilization decision of millet in the whole growth period.

Key words: millet, nitrogen content in canopy, hyperspectral remote sensing, partial least squares regression

摘要：

为构建多品种谷子冠层氮素含量的高光谱监测模型，通过设置不同氮素水平和多品种谷子田间试验，分别获取谷子全生育期冠层高光谱反射率和叶片氮素含量，采用卷积平滑、一阶导数变换对光谱曲线进行预处理，利用光谱数据与氮素含量相关性分析、连续投影算法（successive projections algorithm， SPA）筛选出谷子各生育期和全生育期的氮素敏感波段、植被指数、高光谱特征参数，利用三者组合建立谷子冠层氮素含量的偏最小二乘回归（partial least square regression，PLSR）估测模型，并筛选各生育期最优模型。结果表明，不同生育期最优估测模型有所差异，拔节期基于敏感波段、植被指数、高光谱特征参数的模型精度最高；抽穗期基于敏感波段、植被指数的模型精度最高；灌浆期基于植被指数、高光谱特征参数的模型精度最高；成熟期基于敏感波段、植被指数的模型精度最高；全生育期基于敏感波段与植被指数的模型精度最高（R²=0.903，RPD=3.01）。多输入级综合模型可以充分利用光谱信息，有效提高模型预测精度和稳定性，其中基于敏感波段和植被指数的综合模型表现效果最好，预测集R²均在0.82以上、均方根误差均小于0.119、相对分析误差均大于2.1。以上研究结果为高光谱遥感诊断谷子全生育期氮素盈缺状况与施肥决策提供了理论依据和技术支撑。

关键词: 谷子, 冠层氮素含量, 高光谱遥感, 偏最小二乘回归