基于冠层高光谱特征的油茶叶片碳氮比估算模型构建

A C/N Ratio Estimation Model of Camellia Oleifera Leaves Based on Canopy Hyperspectral Characteristics

叶片碳氮比是反映油茶个体养分利用效率的重要指标,基于冠层高光谱特征估算碳氮比可为油茶养分监测及精准施肥提供重要的科学依据。当前,利用高光谱开展经济林理化性质的研究较少,特别是面向具有花果同期生物学特性的油茶,其高光谱数据除面临共线性问题外,复杂的理化性质也对敏感光谱特征的响应及估算模型构建带来极大挑战。以安徽黄山区域的油茶长林系列品种为研究对象,在野外环境下采集了120株油茶的冠层光谱,选取可见光与近红外谱区400~1000 nm波长范围的高光谱特征进行分析。利用多元散射校正(MSC)和一阶导数(FD)变换对原始高光谱进行处理,并各自构建三种两波段指数(差值指数-DI、比值指数-RI和归一化指数-NDI)。采用相关分析观察不同处理方法下光谱响应特征区域的变化,使用变量组合集群分析(VCPA)提取响应变量并去除共线性得到最优特征变量子集,进一步构建三种机器学习模型(随机森林-RF、支持向量机-SVM和BP神经网络-BPNN)。最后,比较不同处理下光谱参数对模型估算精度的影响,根据模型评价指标得到最优油茶叶片碳氮比估算模型。研究结果表明:(1)经过MSC或FD特征变换的原始光谱协同VCPA能够挖掘更多潜在变量。(2)两波段光谱指数组合扩展了敏感波段的响应区域,进一步增强了VCPA筛选特征变量的能力,FD-RI与FD-NDI处理效果最好。(3)三种机器学习模型整体精度由大到小排序为BPNN>RF>SVM。所有模型中,经过FD-NDI处理的光谱参数构建的BPNN模型预测能力表现最好,训练集和测试集的决定系数(R 2)分别为0.71和0.66,其相对分析误差(RPD)达到1.74。该研究建立了一种收获期油茶叶片碳氮比的最优BPNN估算模型,拓展了油茶高光谱应用的范围。

Leaf C/N ratio is an important indicator reflecting the individual nutrient utilization efficiency of Camellia oleifera.Estimating C/N ratio based on canopy hyperspectral characteristics can provide important theoretical basis for nutrient monitoring and precise fertilization of Camellia oleifera.There are very limited studies on non-timber product forests physical and chemical properties-using hyperspectral data,especially for Camellia oleifera with the synchronous biological characteristics of flowers and fruits.In addition to the collinearity problem,its complex physical and chemical properties pose great challenges to the response of sensitive spectral characteristics and the construction of estimation models.In this study,the Changlin series of Camellia oleifera in the Huanagshan area of Anhui Province was taken as the research objects.The canopy spectra of 120 Camellia oleifera plants were collected in the field,and the hyperspectral characteristics of the 400~1000 nm wavelength range in the visible and near-infrared spectral regions were selected for analysis.Original hyperspectral data were processed by using multiplicative scatter corrections(MSC)and first derivative(FD)transformations,and three types of two-band indices(i.e.,difference index-DI,ratio index-RI,and normalized difference index-NDI)were constructed respectively.Correlation analysis was used to observe the changes inspectral response feature regions under different processing methods.Response variables were extracted by variable combination population analysis(VCPA),and an optimal feature variable subset was obtained by removing collinearity to construct three machine learning models(i.e.,random forest-RF,support vector machine-SVM and BP neural network-BPNN).Finally,the effects of spectral parameters on model estimation accuracy under different treatments were compared,and the optimal estimation model of the C/N ratio of Camellia oleifera leaves was identified according to model evaluation indices.Results showed that:(1)The original spectrum after MSC or FD feature transformation combined with VCPA can mine more potential variables.(2)The combination of a two-band spectral index expands the response region of sensitive bands and further enhances the ability of VCPA to select characteristic variables.FD-RI and FD-NDI are with the best treatment effect.(3)The overall accuracy of the three machine learning models ranked indescending order were BPNN>RF>SVM.Among all models,the BPNN model constructed by FD-NDI spectral parameters has the best prediction ability performance.The determination coefficient(R 2)of the training and test sets are 0.71 and 0.66,respectively,and the relative percent difference(RPD)is 1.74.This study established an optimal BPNN estimation model for the C/N ratio of Camellia oleifera leaves in the harvest period,which expands the application range of hyperspectral of Camellia oleifera leaves.