冷杉叶片氮含量高光谱反演技术研究

Study on hyperspectral Inversion technology ofNitrogen content in Abies fabri leaves

摘要:

氮是植物生长的基本营养元素,对氮含量的监测有助于及时了解植物的代谢过程和健康状况。高光谱技术能够对叶片氮含量进行无损快速的检测,已成为叶片氮含量检测的重要手段之一。采用高光谱技术对冷杉叶片氮含量进行估算,首先分别采用相关系数法、连续投影法、LASSO算法对4种形式的光谱变量R,R′,log(1/R),log(1/R)′进行敏感波段筛选,对筛选后的敏感波段分别采用偏最小二乘、随机森林、支持向量机进行建模反演。结果显示:两种导数形式变量的反演误差最小;且相关系数&Log(1/R)′,LASSO&Log(1/R)′能够有效筛选出叶片氮含量的敏感波段组合,无论采取何种建模方法,其估算效果在几种筛选方法组合中最好,R2＞0.84,RMSE在0.19~0.24,估算效果明显优于全波段建模;随机森林算法对相关系数法和LASS0算法筛选出的Log(1/R)′形式变量的反演误差最小,但对其他形式变量的反演误差变化范围较大,表现并不稳定。3种变量筛选方法筛选出的R形式的变量与前人研究相符,连续投影算法能筛选出更多与叶片其他化学含量相关的波段,证明了3种变量筛选方法的有效性。

Abstract:

Nitrogen is an essential element for plant growth.Monitoring nitrogen content is helpful to understand the metabolic process and health status of plants.Hyperspectral technology,a nondestructive and rapid method to detect leaf nitrogen content,has become one of the important methods for nitrogen analysis.In this paper,hyperspectral technology was used to estimate leaf nitrogen conten in Abies fabri t.Correlation coefficient method,successive projections algorithm and LASSO algorithm were used to select sensitive bands of four spectral variables including R,R′,log(1/R) and log(1/R)′. Partial least square,random forest and support vector machine were used to conduct modeling inversion for the selected sensitive bands.The results showed that the inversion error of the two derivatives was the smallest.The correlation coefficient including Log(1/R)′ and LASSO&Log(1/R)′ could be used effectively to select the combined sensitive band for leaf nitrogen content.The estimation effects using the combination of those two variables gave the best results among several selection methods with R2＞0.84 and RMSE ranged from 0.19~0.24,and the estimation effect was significantly better than that of all the three modeling methods using the whole bands.The inversion errors derived from random forest algorithm in combination with correlation coefficient and Log(1/R)′,and random forest algorithm in combination with LASSO& Log(1/R)′ were smallest.But the error range of the random forest method to other forms of variables was large and its performance was not stable.The variables in the form of R screened by the three variable screening methods were consistent with previous studies.The successive projection algorithm could screen more bands related to other chemical content of leaves,which proved the effectiveness of the three variable screening methods.