文献类型：期刊文献

中文题名：基于地物光谱仪的河流有机碳遥感反演模型

英文题名：Inversion Model for River Organic Carbon Remote Sensing Based on Field Spectroradiometer

作者：刘爽[1,2] 邵学新[1] 孙冠[3] 吴明[1] 焦盛武[1] 张龙[1] 尹园丰[1,4] 蔺星娜[1]

第一作者：刘爽

机构：[1]湿地环境保护与生态修复全国重点实验室,浙江杭州湾湿地生态系统定位观测研究站,中国林业科学研究院亚热带林业研究所,杭州311400;[2]南京林业大学,南京210037;[3]杭州市生态环境局富阳分局,杭州311400;[4]安徽师范大学生态与环境学院,芜湖241002

年份：2025

卷号：5

期号：2

起止页码：23-35

中文期刊名：陆地生态系统与保护学报

外文期刊名：Terrestrial Ecosystem and Conservation

基金：国家自然科学青年基金(D0707-42407359);浙江省基础公益项目(LTGS24C160001);科技基础资源调查专项(2023FY100103)。

语种：中文

中文关键词：河流有机碳;随机森林;支持向量机;地物光谱仪;钱塘江

外文关键词：river organic carbon;random forest;support vector machine;field spectrometer;Qiantang River

分类号：P237

摘要：【目的】对钱塘江干流不同河段的河流有机碳进行采样分析,旨在基于地物光谱仪,构建和评估河流有机碳浓度的最优反演模型。【方法】通过野外采样和实地调查,筛选地物光谱仪获取的钱塘江河流光谱数据,得到河流颗粒有机碳和溶解有机碳的敏感波段,构建了支持向量回归(SVR)、随机森林(RF)、多元线性回归(MLR)和多项式拟合(PolyFit)4种模型,并引入决定系数(R^(2))、残差预测偏差(RPD)、均方根误差(RMSE)、平均绝对误差(MAE)评估模型对河流颗粒有机碳和溶解有机碳的预测精度。【结果】1)河流颗粒有机碳(POC)单波段敏感波段分别为436 nm、456 nm、584 nm、869 nm,Spearman系数为0.14~0.25,呈现弱至中度相关性。通过构建比值性指数,得到双波段敏感波段为RI(667 nm/518 nm)、DI(680 nm/523 nm)、NDI(645 nm/529 nm),显著提升了POC的敏感性。其中DI(680 nm/523 nm)的相关系数最高,为0.73,其次是RI(667 nm/518 nm)和NDI(645 nm/529 nm)。溶解性有机碳(DOC)单波段敏感波段分别为680 nm、867 nm,Spearman系数为0.14~0.18,呈现弱相关性,构建比值性指数后得到双波段敏感波段为RI(584 nm/571 nm)、DI(584 nm/565 nm)、NDI(584 nm/565 nm),DOC双波段组合的敏感性优于单波段,其中DI(584 nm/565 nm)的相关系数最高,为0.67,其次是RI(584 nm/571 nm)和NDI(584 nm/565 nm);2)随着模型变量的增加,各模型的R^(2)、RPD值均呈增大趋势;3)RF-混合波段组合模型对POC和DOC的反演精度均为最优(POC:训练集R^(2)=0.84,RPD=2.49;测试集R^(2)=0.82,RPD=2.41;DOC:训练集R^(2)=0.73,RPD=1.94;测试集R^(2)=0.72,RPD=1.91)。【结论】随机森林-混合波段组合模型在钱塘江干流的颗粒有机碳和溶解有机碳浓度的预测中均表现优异。证实了高光谱数据在河流碳动态监测的可行性,为区域河流有机碳浓度的长期监测提供了可靠的技术手段,进而为揭示流域-河口连续体中有机碳的迁移路径提供了科学参考和方法指导。
【Objective】This study conducted sampling and analysis of riverine organic carbon in different sections of the Qiantang River mainstream,aiming to construct and evaluate the optimal inversion model for river organic carbon concentration based on geophysical spectrometer.【Method】Through field sampling and surveys,river spectral data from the Qiantang River acquired by the spectrometer were screened to identify sensitive bands for particulate organic carbon(POC)and dissolved organic carbon(DOC).Four models were constructed:Support Vector Regression(SVR),Random Forest(RF),Multiple Linear Regression(MLR),and Polynomial Fitting(PolyFit).Model performance was assessed using the coefficient of determination(R^(2)),residual prediction deviation(RPD),root mean square error(RMSE),and mean absolute error(MAE).【Result】1)For POC,single-band sensitive wavelengths were 436 nm,456 nm,584 nm,and 869 nm,with Spearman coefficients of 0.14~0.25(weak to moderate correlations).Ratio-based indices significantly improved sensitivity,with dual-band combinations RI(667 nm/518 nm),DI(680 nm/523 nm),and NDI(645 nm/529 nm)showing the highest correlations.DI(680 nm/523 nm)achieved the highest correlation coefficient(0.73),followed by RI and NDI.For DOC,single-band sensitive wavelengths were 680 nm and 867 nm(Spearman coefficients:0.14-0.18,weak correlations).Dual-band indices RI(584 nm/571 nm),DI(584 nm/565 nm),and NDI(584 nm/565 nm)outperformed single bands,with DI(584 nm/565 nm)exhibiting the highest correlation(0.67).2)As model variables increased,R^(2) and RPD values improved across all models.3)The RF model with hybrid band combinations achieved optimal accuracy for both POC(training set:R^(2)=0.84,RPD=2.49;testing set:R^(2)=0.82,RPD=2.41)and DOC(training set:R^(2)=0.73,RPD=1.94;testing set:R^(2)=0.72,RPD=1.91).【Conclusion】The Random Forest hybrid band combination model demonstrated superior performance in predicting POC and DOC concentrations.This study validates the feasibility of hyperspectral data for monitoring river carbon dynamics and provides a reliable technical approach for long-term monitoring of regional river organic carbon concentrations.The findings offer scientific insights into the transport pathways of organic carbon in the river-estuary continuum.