基于光谱敏感变量优选的澳洲坚果叶片氮素含量估算

Estimation for Nitrogen Content of Macadamia Leaves Based on Optimized Spectral Sensitive Variables

利用高光谱遥感技术探索澳洲坚果叶片氮素含量估算方法,以实现澳洲坚果氮素营养快速诊断。本研究以临沧和西双版纳为研究区,获取澳洲坚果品种O.C和HAES344叶片的光谱反射率和氮素含量,首先采用对数变换、导数变换及其组合对原始光谱反射率进行多种数学变换,然后分析澳洲坚果叶片氮素含量与不同变换形式光谱数据的相关性;在决定系数较大的原则下,选择决定系数曲线图中波峰特征点对应的波长作为氮素敏感波长,从而得到相应的氮素敏感光谱变量;运用逐步回归法对氮素敏感光谱变量进一步优化,并采用多元线性回归(MLR)、偏最小二乘回归(PLSR)和支持向量回归(SVR)3种方法构建澳洲坚果叶片氮素含量估算模型;最后,分别利用验证集和测试集对构建的澳洲坚果叶片氮素含量估算模型性能进行测试。结果显示,MLR、PLSR、SVR等3种模型估算能力均表现良好,验证集和测试集的相对分析误差(RPD)均在2.0以上;其中,PLSR模型为最优估算模型,验证集和测试集的RPD分别为2.099和2.110。从反射率(R)、对数变换(LR)、一阶导数(FDR)、对数变换的一阶导数(FDLR)、二阶导数(SDR)、对数变换的二阶导数(SDLR)等6种变换光谱数据中优选的19个氮素敏感光谱变量,对氮素光谱响应具有较强的稳定性;基于优选的19个氮素敏感光谱变量,用常规的回归建模方法均能取得良好的估算效果,且具有较强的区域普适性。本研究从多种变换光谱数据中优选氮素敏感光谱变量,为澳洲坚果叶片氮素含量估算提供新思路。

Hyper-spectral remote sensing technology was used to explore the estimation method of nitrogen content in the leaves of macadamia to achieve a rapid diagnosis of nitrogen nutrition in macadamia trees.Lincang and Xishuangbanna were chosen as the research area to obtain the spectral reflectance and nitrogen content of the leaves of macadamia varieties O.C and HAES344.Firstly,multiple mathematical transformations were performed on the original spectral reflectance using logarithmic transformation,derivative transformation,and their combinations.Then,the correlation between nitrogen content of macadamia leaves and spectral data of different transformation forms was analyzed.Under the principle of larger determination coefficient,the wavelength corresponding to the peak characteristic point in the determination coefficient curve was selected as the nitrogen sensitive wavelength,thus the corresponding spectral variables of nitrogen sensitivity were obtained.Stepwise regression was used to further optimize the nitrogen sensitive spectral variables,and the methods of multiple linear regression(MLR),partial least squares regression(PLSR),and support vector regression(SVR)were used to construct the nitrogen content estimation models for macadamia leaves.Finally,the performance of the models was tested using validation and test sets,respectively.The results showed that the MLR,PLSR and SVR models all performed well in estimation,and the ratio of performance to standard deviate(RPD)of both the validation and test sets were above 2.0.Among them,the PLSR was the optimal estimation model,its RPD of the validation set and the test set was 2.099 and 2.110,respectively.The 19 nitrogen sensitive spectral variables selected from 6 types of transformation spectral data,including reflectance(R),logarithmic transformation of reflectance(LR),first derivative of reflectance(FDR),first derivative of logarithmic transformation of reflectance(FDLR),second derivative of reflectance(SDR),and second derivative of logarithmic transformation of reflectance(SDLR)had strong stability in nitrogen spectral response.Based on the selected 19 nitrogen sensitive spectral variables,the conventional regression modeling methods could achieve good estimation results and had strong regional universality.In this study,nitrogen sensitive spectral variables were selected from a variety of transform spectral data,which provided a new idea for the nitrogen content estimation of macadamia leaves.