基于近红外光谱技术和混合学习框架的橡胶树叶片氮含量估算

Estimating Nitrogen Concentration of Rubber Leaves Based on a Hybrid Learning Framework and Near-Infrared Spectroscopy

叶片氮含量(LNC)是判断橡胶树营养状态的一个重要量化指标,快速准确地检测橡胶树的叶片氮含量对于保证橡胶树的生长和天然橡胶的产量是非常十分必要的。利用近红外光谱技术对119片橡胶叶片的叶面氮含量进行了定量分析,建立了高精度的预测模型,实现了对橡胶叶片氮含量的快速精准检测。采集海南橡胶叶作物实验对象,首先使用GaiaField-F-N17E光谱仪测量橡胶叶片的近红外光反射率数据,波长范围为942~1680 nm。然后,消除光谱数据中的异常样本,分别使用了三种不同的预处理方法对数据进行处理并比较它们对模型精度的提升效果。由于橡胶叶片的近红外光谱数据存在着大量的冗余信息和高度共线的光谱特征波段,因此,提出了一种基于改进后的模因框架(IMF)的结合竞争自适应重加权采样(CARS)和近邻搜索(NNS)的混合变量选择方法,采用该算法消除光谱中的冗余信息并进行二次优化,从全波段中提取28个作为建模波段。最后,使用偏最小二乘回归(PLSR)和最终选取的波段建立橡胶叶片的LNC估算模型。为了验证所提方法的优越性,进一步使用CARS,连续投影(SPA)和传统模因算法(MA)的变量选择算法建立模型作为对比。结果表明,多元防散射效正(MSC)处理后的光谱曲线和基于IMF框架的CARS-NNS算法所建立的模型在预测集上的表现最佳:均方根误差(RMSEp)达到0.116,决定系数(R_(p)^(2))为0.951,两项评价指标均优于其他的预测模型。综上所述,基于近红外光谱技术和使用混合学习IMF框架构建的预测模型能够很好地揭示光谱数据与橡胶树叶片氮含量两者之间的关系,可为橡胶林的养分诊断提供必要的技术支持,保证橡胶树的良好生长,以提升天然橡胶的产量和质量。

Leaf nitrogen Concentration(LNC)is an important criterion for determining the nutritional status of rubber trees.Rapid and accurate detection of rubber trees’leaf nitrogen content is necessary to ensure the growth of rubber trees.In this paper,the leaf nitrogen content of 119 rubber leaves was quantitatively analyzed by near-infrared spectroscopy,a high-precision prediction model was established,and the rapid and accurate detection of nitrogen content in rubber leaves was realized.The experimental objects of rubber leaf crops in Hainan were collected.First,the GaiaField-F-N17E spectrometer was used to measure rubber leaves’near-infrared spectral reflectance data,with a wavelength range of 942nm to 1680nm.Then,abnormal samples in the measured spectral data were eliminated,and three different preprocessing methods were used to transform the data and compare their effects on improving model accuracy.Due to the massive redundant information and highly collinear spectral feature bands in the near-infrared spectral data of rubber leaves,a hybrid variable selection algorithm consisting of machine learning and evolutionary algorithms was proposed.it can effectively eliminate the redundancy and collinearity of spectral features and use the proposed method to extract the 28 bands from all 224 spectral bands effectively.Finally,using partial least squares regression(PLSR)and the selected spectral bands were used to establish the LNC estimation model of the rubber leaves.The results show that the spectral curve after multivariate anti-scattering effect(MSC)processing and the estimation model established by the CARS-NNS algorithm performed on the prediction set as follows:the RMSE p reaches 0.116,and the coefficient of determination is 0.116.R_(p)^(2) was 0.951.Both evaluation metrics were better than other models.In conclusion,the prediction model based on NIR spectroscopy and the hybrid learning IMF framework can well reveal the relationship between the spectral data and the nitrogen content of rubber leaves,providing a necessary technology for the nutrient diagnosis of rubber forests.Ensure the good growth of rubber trees to improve the yield and quality of natural rubber.