模型约束与机器学习下的植物类胡萝卜素和叶绿素含量反演方法  

Retrieval of Plant Carotenoids and Chlorophyll Contents With Model Constraints and Machine Learning

作　　者：汤馥睿 徐媛媛[1] 耿芫 蔡顾斌 杨帆[1] 李雨晨 季颖[1] TANG Fu-rui;XU Yuan-yuan;GENG Yan;CAI Gu-bin;YANG Fan;LI Yu-chen;JI Ying(College of Physics and Electronic Engineering,Jiangsu University,Zhenjiang 212000,China)

机构地区：[1]江苏大学物理与电子工程学院,江苏镇江212000

出　　处：《光谱学与光谱分析》2024年第8期2174-2182,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(11874184)资助。

摘　　要：叶绿素和类胡萝卜素含量是评价植物健康状况的一个重要指标。PROSPECT模型与机器学习耦合反演植被生化特性已得到广泛应用。但由于叶片方向半球反射率因子(DHRF)光谱和二向反射率因子(BRF)光谱之间的差异,耦合模型的应用范围受到限制。为此,以北美地区植物叶片光谱数据库(EcoSIS)作为实验数据集,提出PROSPECT模型作为机器学习的附加约束形成混合数据集,对此混合数据集利用连续小波变换(CWT)产生的小波系数谱和一阶导(FD)产生谱,提出三种全光谱域和VNIR光谱子域下的植物叶片叶绿素、类胡萝卜素的光谱特征变量筛选策略,即是:竞争性自适应重加权算法(CARS)、连续投影算法(SPA)和主成分分析法(PCA)。由此,基于上述2×2×3=12种不同光谱处理方法、特征提取方法组合,分别建立了植物叶片叶绿素和类胡萝卜素含量的人工神经网络(ANN)预测模型。进而开展了不同模型下的预测精度对比分析,结果表明:PROSPECT模型约束下的模拟数据一定程度增强了机器学习的训练集质量;经一阶导、小波变换处理的光谱能较好地减少DHRF模拟光谱和BRF实测光谱间的偏差,并在结合特征提取算法CARS后进一步提升了预测表现。在全光谱域下的FD+CARS组合对叶片叶绿素的反演效果最佳,测试集R2为0.8064,RMSE为2.9114;在VNIR光谱子域下的CWT+CARS组合对叶片类胡萝卜素最佳,测试集R2为0.7972,RMSE为0.4141。该方法可为研究人员从叶片BRF光谱及其他近端反射率图像更精确、高效地提取植物叶片生化特征提供参考。The chlorophyll and carotenoid content is an important indicator for evaluating the health status of plants.The PROSPECT model,coupled with machine learning,has been widely used to retrieve the biochemical properties of vegetation.However,the application of the coupled model is limited due to the differences between the leaf-directional hemispherical reflectance factor(DHRF)spectra and the bidirectional reflectance factor(BRF)spectra.This paper utilizes the leaf spectral database of North American plant(EcoSIS)as the experimental dataset and introduces the PROSPECT model as an additional constraint for machine learning.This approach creates a hybrid dataset by employing wavelet continuous wavelet transform(CWT)to generate the wavelet coefficient spectrum and the derivative spectrum generated by the first-order derivative(FD).Three kinds of feature extraction algorithms,namely competitive adaptive reweighting algorithm(CARS),successive projection algorithm(SPA),and principal component analysis(PCA)were applied to extract spectral features for chlorophylls and carotenoids in the full-spectral domains and the subdomain of VNIR spectroscopy.Based on the above 12 combinations of different methods,artificial neural network(ANN)prediction models for chlorophyll and carotenoids were separately established.The results show that the simulated data under the constraint of the PROSPECT model enhanced the quality of the training set for machine learning to a certain extent.Additionally,the spectra processed by the first-order derivatives and wavelet transforms were able to reduce better the bias between the simulated spectra of the DHRF and the measured spectra of the BRF.The best inversion of leaf chlorophyll is achieved with the FD+CARS combination in the whole spectral domain,yielding a test set R2 of 0.8064 and RMSE of 2.9114.Meanwhile,the CWT+CARS combination in the VNIR spectral sub-domain offers the best results for leaf carotenoids,with a test set R2 of 0.7972 and RMSE of 0.4141.The proposed method can provide researc

关 键 词：光谱数据 模型约束 小波系数谱 一阶导数谱 叶绿素 类胡萝卜素 精准反演 

分 类 号：O433.4[机械工程—光学工程]