摘要
油膜厚度是水面溢油量估算的一个关键参数。为了确定高光谱数据探测水面油膜厚度的可行性,在实验室内,以原油作为实验油品,以石英卤素灯模拟太阳光源,以ASD FieldSpec3作为光谱探测仪器,开展了不同厚度油膜模拟及其反射率光谱测量实验,获取油膜厚度-反射率光谱数据27组。为了充分利用光谱信息,选择偏最小二乘法(partial least squares,PLS)进行油膜厚度-光谱反射率建模,样本数据中21组用于建模,6组用于验证。研究结果表明,当主成分分量个数为5时,PLS模型具有最佳效果,5个主成分分量累积解释了74%的自变量信息和99.8%的因变量信息,模型的预测能力达到92.8%,建模样本和验证样本的均方根误差(root mean squared error,RMSE)分别为0.01和0.04,说明所建立的PLS模型具有较好的预测能力和稳定性。通过与传统曲线拟合模型的对比,PLS模型在误差方面无论是建模样本还是验证样本均优于传统的经验模型,因而认为基于PLS模型可以实现高光谱数据水面油膜厚度估算。
Thickness of oil slick is an important parameter of oil spill volume.In order to confirm the feasibility of oil thickness estimation with hyperspectral data,the authors used ASD FieldSpec3,quartz halogen lamp and crude oil for a laboratory experiment which simulates oil slick and spectral measurement.27 pairs of oil thickness and reflection data were acquired.To make full use of spectral information of the hyperspectral data,the authors selected partial least square (PLS) to slick thickness and reflection modeling with 21 set model data and 6 test data set.Model result shows that PLS model expresses optimal effect when five principal components are selected which interpret 74% information of independent variables and 99.8% information of dependent variable,the prediction capability of the model runs up to 92.8%.The root mean squared error is 0.01 for modeling samples and 0.04 for validation samples.The PLS model shows better accuracy of modeling and validation error compared with traditional model,and thus it can be used in oil slicks thickness modeling with hyperspectral data.
作者
邢学文
刘松
许德刚
钱凯俊
XING Xuewen;LIU Song;XU Degang;QIAN Kaijun(Petrochina Research Institute of Petroleum Exploration and Development,Beijing 100083,China;ChinesePetroleum Safety and Environmental Protection Technology Research Institute,Beijing 102206,China)
出处
《国土资源遥感》
CSCD
北大核心
2019年第2期111-117,共7页
Remote Sensing for Land & Resources
基金
中国石油“十三五”重大科技项目“南海油气形成条件与勘探技术研究及重大目标优选”(编号:2016A-1003)资助
关键词
油膜实验
油膜厚度
高光谱数据
偏最小二乘法
oil slick simulation experiment
thickness of oil slicks
hyperspectral data
partial least square