基于多光谱图像的幼龄白木香含锌量无损诊断

通过探讨白木香冠层光谱和形状特征与叶片含锌量的模型关系,实现幼龄白木香冠层含锌量的快速无损诊断,为实现白木香智能化培育经营提供新思路。以幼龄白木香为研究对象,通过多光谱相机获取白木香冠层图像,结合相位相关法及贝叶斯分割法精确提取白木香冠层,在应用偏最小二乘（partial least squares, PLS）算法对图像光谱和形状特征进行降维的基础上,分别构建偏最小二乘回归（partial least squares regression, PLSR）模型和偏最小二乘-广义可加模型（partial least squares- generalized additive models, PLS-GAM）以图像特征对含锌量进行估测和分析,并通过比较模型评价指标与常用套索回归（lasso regression, LassoR）和多元逐步回归（multiple stepwise regression, MSR）模型进行对比,确定适用于白木香锌含量估测的最佳模型。研究表明：（1）结合相位相关法和贝叶斯算法能够较好地分割出白木香冠层图像,效果显著优于对各波段图像进行直接分割的方法;（2）基于多光谱图像特征提取6个主成分CF₁、CF₂、CF₃、CF₄、CF₅和CF₆,PLSR建模分析结果表明CF₁和CF₂与白木香冠层含锌量具有显著的线性关系,模型调整后R²_adj为0.475;（3）PLS-GAM建模分析结果表明,CF₁、CF₂和CF₄与白木香冠层含锌量均存在显著的非线性关系,模型调整后R²_adj为0.679,显著高于基于线性关系构建的PLSR模型;（3）经过模型评价对比,PLS-GAM模型估测精度最高,RMSE为0.095,较PLSR、LassoR、MSR模型分别降低了26.4%、43.1%和34.9%,为适用于估测白木香冠层含锌量的最优模型。因此,结合相位相关法及贝叶斯分割法能够实现对白木香冠层多光谱图像的精准分割,基于光谱和形状特征构建的PLS-GAM模型对白木香冠层含锌量具有良好的估测效果,有利于推动白木香微量元素诊断的研究进程,对幼龄白木香的智能化作业有重要意义。

Nondestructive Diagnosis of Zinc Content in Young Seedlings of Aquilaria sinensis (Lour.) Spreng. Based on Multispectral Image

By discussing the relationship between Aquilaria sinensis (Lour.) Spreng. canopy spectrum and shape characteristics and leaf zinc content, the rapid and non-destructive diagnosis of zinc content in young seedlings of A. sinensis canopy could be realized, which would provide a new idea for the realization of intelligent cultivation and management the seedlings of A. sinensis. In this study, the canopy images of young seedlings of A. sinensis were obtained by multi-spectral camera, and the canopy was extracted accurately by phase correlation method and bayesian segmentation method. Then, based on the partial least squares (PLS) algorithm for dimensionality reduction of image spectrum and shape features, partial least squares regression (PLSR) model and partial least squares-generalized additive model (PLS-GAM) were constructed to estimate and analyze zinc content. Finally, PLSR model and PLS-GAM were compared with lasso regression (LassoR) model and multiple stepwise regression (MSR) model to determine the best model for estimating zinc content of the seedlings of A. sinensis based on the model evaluation indexes. The phase correlation and Bayesian algorithm could segment the A. sinensis seedlings canopy image, which is better than the method of direct segmentation for each band image. Six principal components CF₁, CF₂, CF₃, CF₄, CF₅ and CF₆ were extracted based on multi-spectral image features. The analysis results of PLSR model showed that CF₁ and CF₂ had a noticeable linear relationship with zinc content in A. sinensis seedlings canopy, and the R²_adj of model was 0.475. The analysis results of PLS-GAM showed that CF₁, CF₂ and CF₄ had noticeable nonlinear relationship with zinc content in A. sinensis seedlings canopy, and the R²_adj of the model was 0.679, which was significantly higher than that of the PLSR model based on linear relationship. After the evaluation and comparison of models, PLS-GAM had the highest estimation accuracy and the RMSE was 0.095, which was 26.4%, 43.1% and 34.9% lower than PLSR model, LassoR model and MSR model, respectively. It was the optimal model for estimating zinc content in A. sinensis seedlings canopy. Consequently, the combination of phase correlation method and Bayesian segmentation method can achieve accurate segmentation of multi-spectral images of A. sinensis seedlings, and the PLS-GAM model based on spectrum and shape features has good estimation effect on zinc content of A. sinensis canopy. It is conducive to promoting the research process of the diagnosis of trace elements in A. sinensis, and it is of great significance to the precise operation of A. sinensis seedlings.