基于微分变换的湿地植物高光谱全氮反演  

Hyperspectral Total Nitrogen Inversion of Wetland Plants Based on Differential Transformation

作　　者：黎舟 杨思林[1,2] 刘云根[1,2] 王妍[1,2] 张超 肖羽芯 徐红枫 LI Zhou;YANG Silin;LIU Yungen;WANG Yan;ZHANG Chao;XIAO Yuxin;XU Hongfeng(College of Ecology and Environment,Southwest Forestry University,Kunming 650224,China;Yunnan Provincial Key Laboratory of Ecological Environment Change and Pollution Control in Mountainous and Rural Areas,Southwest Forestry University,Kunming 650224,China;College of Forestry,Southwest Forestry University,Kunming 650224,China)

机构地区：[1]西南林业大学生态与环境学院,云南昆明650224 [2]西南林业大学,云南省山地农村生态环境演变与污染治理重点实验室,云南昆明650224 [3]西南林业大学林学院,云南昆明650224

出　　处：《环境科学研究》2022年第5期1268-1276,共9页Research of Environmental Sciences

基　　金：国家自然科学基金项目(No.32160405,21767027)。

摘　　要：叶片全氮含量是监测和诊断湿地植物生理状况及生长趋势的重要指标,利用高光谱技术监测湿地植物氮含量,对于理解湿地生态系统氮循环具有重要意义.为探究湿地植物叶片全氮含量遥感光谱的估算方法,以云南省大理西湖湿地公园优势植物芦苇(Phragmites australis)和茭草(Zizania caduciflora)全氮含量为研究对象,对叶片光谱数据进行预处理并建立二者的关系模型,包括单变量模型、多变量模型(偏最小二乘回归模型和BP神经网络模型),并利用决定系数(r^(2))和均方根误差(RMSE)对模型精度进行检验.结果表明:①不同形式的光谱变换增强了植物全氮含量与光谱变量的细节特征,二者的短波红外波段相关性强于可见光近红外波段.芦苇二阶微分(R")反射率与全氮含量在1682 nm处相关性最强,相关系数为0.70;茭草平方根二阶微分[(√R)"]反射率与全氮含量在1190 nm处相关性最强,相关系数高达-0.80.②不同植物类型相比,利用茭草的变换光谱反射率建立的单变量和偏最小二乘回归模型建模精度都高于芦苇.③不同回归模型相比,BP神经网络模型的精度最高,其芦苇和茭草全氮含量估算模型的r^(2)均为0.96,均方根误差(RMSE)分别为0.63、0.47,是建立湿地植物光谱与全氮含量关系的最优模型.研究显示,BP神经网络模型对湿地植物氮含量的预测精度较高,且计算速度快,不仅可为人工智能技术在湿地监测与管理提供有力的科学依据,而且可以为湖泊水环境污染治理应用提供新思路.Total nitrogen content(TN)in leaves is an important indicator for monitoring and identifying the physiological status and growth trend of wetland plants.Using hyperspectral remote sensing technology to estimate wetland plant leaf nitrogen content is essential for understanding the nitrogen cycle of wetland ecosystems.This study aimed to investigate the potential of leaf spectra of wetland vegetation in estimating the nitrogen content.The TN content of leaves of typical wetland plants Phragmites australis and Zizania caduciflora in Xihu Lake wetland park of Yunnan Province was studied.The plant leaf reflectance spectra of sample sites were acquired with all ASD Field Spec 3 spectrometer(350-2500 nm),and leaf total nitrogen contents were determined by Kjeldahl nitrogen measurement method after acquiring the leaf reflectance spectra.Ten advanced differential transformation spectral algorithms were used for spectral pre-treatments.The correlations between leaf nitrogen contents and the differential transformation spectrum were evaluated by partial correlation analysis,and then univariate models and multivariate models including partial least squares regression and back propagation(BP)neural network algorithm model were established.Moreover,the accuracy of all the models was tested through Pearson correlation coefficient of determination(r^(2))and root mean square error(RMSE).The results showed that:(1)The differential transformation can effectively improve the sensitivity of the original spectrum leaf nitrogen content inversion,and fully reflect the sensitivity of short-wave infrared wave band representing leaf TN content.The correlation between second-order differential(R")reflectance of Phragmites australis and TN content was the highest at 1682 nm,with the correlation coefficient of 0.70.The correlation between square root second-order differential((√R)")reflectance of Zizania caduciflora and TN content was the highest at 1190 nm,with the correlation coefficient of-0.80.(2)Comparing different types of wetland pl

关 键 词：光谱变换 单变量模型 偏最小二乘回归 BP神经网络 

分 类 号：X503.23[环境科学与工程—环境工程] Q948[生物学—植物学]