融合光谱反射率与时频特征的坡垒叶绿素含量估测模型研究  

Estimation of chlorophyll content in Hopea hainanensis based on fusedspectral reflectance and time-frequency features

作　　者：袁莹 王雪峰[1] 石蒙蒙 王鹏 YUAN Ying;WANG Xuefeng;SHI Mengmeng;WANG Peng(Research Institute of Forest Resource Information Techniques,Chinese Academy of Forestry,Key Laboratory of Forest Management and Growth Simulation of National Forestry and Grassland Administration,Beijing 100091,China)

机构地区：[1]中国林业科学研究院资源信息研究所,国家林业和草原局森林经营与生长模拟重点实验室,北京100091

出　　处：《西北农林科技大学学报（自然科学版）》2023年第11期97-105,共9页Journal of Northwest A&F University(Natural Science Edition)

基　　金：国家自然科学基金项目(32071761)。

摘　　要：【目的】研究估测坡垒叶绿素含量的高光谱模型,探讨融合光谱反射率与时频特征估测叶绿素含量的可行性,为坡垒叶绿素含量的快速无损检测提供参考。【方法】以2年生坡垒为研究对象,设置N0(0 kg/hm~2)、N1(35 kg/hm~2)、N2(70 kg/hm~2)3个氮肥水平,测定不同氮肥水平下坡垒冠层叶绿素含量和光谱曲线,采用离散傅里叶变换(DFT)和短时傅里叶变换(STFT)方法提取光谱时频特征,采用偏最小二乘回归(PLSR)算法,构建基于光谱反射率、时频特征及其二者融合特征的不同氮水平的分区模型以及全氮肥水平下的综合模型,估测不同氮肥水平下坡垒叶绿素含量,并以决定系数(R~2)、平均绝对百分比误差(MAPE)、相对均方根误差(RRMSE)作为评价指标比较各模型精度,确定最优估测模型。【结果】(1)不同叶绿素含量下坡垒冠层光谱反射特征的整体变化趋势相似,在可见光波段(380~750 nm)内,光谱反射率随着叶绿素含量的增加而降低。(2)在构建的分区模型中,基于光谱反射率和时频特征构建的模型均能较好地估测坡垒叶绿素含量,检验R~2分别为0.626~0.816,0.662~0.797,检验MAPE分别为4.966%~9.269%,6.029%~8.181%,检验RRMSE分别为6.827%~11.593%,8.247%~9.792%;基于融合光谱反射率与时频特征(融合特征)构建的分区模型检验R~2为0.913~0.951,检验MAPE和RRMSE均低于10%。(3)在构建的综合模型中,当引入哑变量时,基于融合光谱反射率与时频特征(融合特征)构建模型的检验R~2为0.814,MAPE和RRMSE分别为7.212%和8.578%,二者较基于光谱反射率构建的模型分别降低了37%和36%,较基于时频特征构建的模型分别降低了47%和45%。【结论】基于融合光谱反射率和时频特征(融合特征)构建的模型能够提高坡垒叶绿素含量估测的精度,对于不同氮肥水平下坡垒叶绿素含量有较好的估测效果,可作为估测坡垒叶绿素含量的优选方法。【Objective】The hyperspectral method for estimating chlorophyll content in Hopea hainanensis was established and the feasibility of combining spectral reflectance and time-frequency features was discussed to provide references for rapid and nondestructive monitoring.【Method】Two-year-old Hopea hainanensis was selected for obtaining canopy chlorophyll content(SPAD value)and spectral curves under three nitrogen fertilizer levels of N0(0 kg/hm 2),N1(35 kg/hm 2)and N2(70 kg/hm 2).The spectral time-frequency features were extracted by the discrete Fourier transform(DFT)and short time Fourier transform(STFT).Then,the partial least squares regression(PLSR)models were constructed to estimate chlorophyll contents under different nitrogen management levels based on spectral reflectance,time-frequency features and their fusion features.Finally,R 2,mean absolute percentage error(MAPE)and relative root mean square error(RRMSE)were used to evaluate model accuracy and determine the optimal estimation model.【Result】(1)The spectral reflectance decreased with the increase of chlorophyll content in the visible band(380-750 nm)by comparing the spectral reflectance of the minimum,near average and maximum values of chlorophyll.(2)In the zoning models constructed under single nitrogen levels,the spectral reflectance and time-frequency features obtained good estimation results,with R 2 of 0.626-0.816 and 0.662-0.797,MAPE of 4.966%-9.269%and 6.029%-8.181%and RRMSE of 6.827%-11.593%and 8.247%-9.792%,respectively.The R 2 of fusion feature models reached 0.913-0.951 and MAPE and RRMSE were both lower than 10%.(3)In the comprehensive models constructed at the full levels after introducing dummy variables,R 2 of the fused model was 0.814,and MAPE and RRMSE were 7.212%and 8.578%,which were 37%and 36%lower than the models based on spectral reflectance and 47%and 45%lower than the models based on time-frequency features.【Conclusion】The fusion of spectral reflectance and time-frequency features effectively improved the accuracy of H

关 键 词：坡垒 叶绿素含量 高光谱 时频特征 短时傅里叶变换 

分 类 号：S796[农业科学—林木遗传育种]