优化光谱指数建立水稻叶片SPAD的高光谱反演模型

利用高光谱反射率光谱的特征波段构建光谱指数，建立叶绿素含量反演模型是实现水稻生产精准调控和科学管理的必要手段之一。为了建立适用于拔节孕穗期水稻叶片叶绿素相对含量（SPAD）的高光谱反演模型，分别获取了拔节孕穗期水稻叶片的高光谱和SPAD数据，利用小波分析法对原始光谱反射率曲线进行降噪处理，并对基于积分运算的光谱指数NAOC进行简化，获得了基于双波段简化运算的优化光谱指数。利用相关分析法计算由原始反射率光谱R和数学变换光谱LgR、1/R和R构建的优化光谱和变换光谱指数与水稻叶片SPAD的相关系数，获得了以积分限（a，b）为横、纵坐标的相关系数二维矩阵，并绘制相关性等势图，得到相关系数最高的3个波段组合：R（641，790）（0.872 6），R（653，767）（0.871 7）和R（644，774）（0.871 6），计算出20个原始样本中3个积分波段组合所对应的60个优化光谱指数值，按照2∶1的比例划分为建模集和验证集，建立了三种水稻叶片SPAD反演模型：偏最小二乘回归（PLSR）、支持向量机（SVM）和BP神经网络模型。结果显示：利用优化光谱和变换光谱指数建立的3种水稻叶片SPAD反演模型决定系数R2均大于0.79，归一化均方根误差NRMSE则小于5.4%。其中BP神经网络相对于其他两种模型具有较高的拟合度，预测精度也相对较高，建模集R2=0.842 6，NRMSE=5.152 7%；验证集R2=0.857，NRMSE=4.829 9%。总体来看，基于双波段简化运算后的优化光谱和变换光谱指数建立拔节孕穗期水稻叶片SPAD反演模型是可行的；对比分析3种模型反演结果发现，BP神经网络对水稻叶片SPAD的反演效果较好。该工作对提高拔节孕穗期水稻精准调控技术和建立水稻生产的科学管理体系具有一定的参考价值。

Hyperspectral Inversion Model for SPAD of Rice Leaves Based on Optimized Spectral Index

It is one of the necessary measures to achieve accurate regulation and scientific management of rice production using the characteristic bands of hyperspectral reflectance curve to construct spectral index and establish chlorophyll content inversion model.In order to establish a hyperspectral inversion model for relative chlorophyll content (SPAD values) of rice leaves at the jointing and booting stage, the hyperspectral data and SPAD values of rice leaves at the jointing and booting stage were obtained respectively. The original spectral reflectance curve was denoised utilizing using the wavelet analysis method, and the spectral index NAOC based on the integral operation was simplified to obtain a simple spectral reflectance curve based on dual-wavelength. The correlation coefficients between SPAD values of rice leaves at jointing and booting stage and the optimized spectral and transformed spectral indices constructed by the original reflectance spectrum R and mathematical transformation spectrum LgR, 1/R and R were calculated by the correlation analysis method. The two-dimensional matrix of correlation coefficients with the integration limit (a, b) as the abscissa and ordinate was obtained. Three band combinations with the highest correlation coefficient: R (641, 790) (0.872 6), R(653, 747) (0.871 7) and R (644, 774) (0.871 6) were selected to calculate 60 optimized spectral indices corresponding to the combination of three integral bands in 20 original samples, which were divided into modeling set and validation set according to the ratio of 2∶1. Three SPAD inversion models of rice leaves were established: partial least squares regression model (PLSR), support vector machine (SVM) and BP neural network. The results showed that: the determination coefficients of the three SPAD inversion models were all greater than 0.79, and the normalizedroot mean square error was less than 5.4%. Compared with the other two models, BP neural network has the highest fitting degree and the highest prediction accuracy, the modeling set R2=0.842 6, NRMSE=5.152 7%; the verification set R2=0.857, NRMSE=4.829 9%. In general, it is feasible to establish an SPAD inversion model of rice leaves at the jointing and booting stage based on optimized spectrum and transformed spectrum index after simplified operation of dual-wavelength. The results of SPAD inversion of rice leaves by BP neural network are ideal and better than the other two inversion models, which have a certain reference value for improving the precision control technology of rice at jointing and booting stage establishing a scientific management system for rice production.