Ultraviolet-induced fluorescence of oil spill recognition using a semi-supervised algorithm based on thickness and mixing proportion-emission matrices.
Ultraviolet-induced fluorescence of oil spill recognition using a semi-supervised algorithm based on thickness and mixing proportion-emission matrices.
复制标题
DOI：
10.1039/d2ay01776h
复制
发表时间：
2023-03
期刊：
影响因子：
--
通讯作者：
Bowen Gong;Hong-Ji Zhang;Xiaodong Wang;Ke Lian;Xinkai Li;Bo Chen;Hanlin Wang;Xiaoqian Niu
中科院分区：
文献类型：
--
作者：
Bowen Gong;Hong-Ji Zhang;Xiaodong Wang;Ke Lian;Xinkai Li;Bo Chen;Hanlin Wang;Xiaoqian Niu
近年来，海上溢油事故在海上石油开采和运输过程中频繁发生，严重破坏了生态平衡。溢油的准确监测在估计溢油量、确定责任和清理方面起着至关重要的作用。泄漏到自然环境中的油并不是单一种类的油，而是多种油品的混合物，并且在风浪的影响下，海面上的油膜厚度不均匀。提出了增加混合油膜厚度维数和混合比例维数来减弱检测环境对荧光测量结果的影响。为了保持不同油膜厚度和不同混合比例油膜数据之间的关系，在实验室测量了海水表面混合油膜的三维荧光光谱数据，建立了厚度-荧光矩阵和比例-荧光矩阵。根据荧光激光雷达方程，研究了荧光光谱的非线性变化。该工作通过求和归一化和二维主成分分析（2DPCA）对数据进行预处理，并将降维结果作为两个特征点视图。然后，半监督分类的协同训练（co-training）与K-近邻（KNN）和决策树（DT）被用来识别样本。结果表明，该耦合模型的平均总体准确度可达100%，比仅厚度视图的准确度高出20.49%。利用未标记数据可以降低数据采集成本，提高分类精度和泛化能力，为海洋环境中光谱相似油种的判别提供理论意义和应用前景。
In recent years, marine oil spill accidents have been occurring frequently during extraction and transportation, and seriously damage the ecological balance. Accurate monitoring of oil spills plays a vital role in estimating oil spill volume, determination of liability, and clean-up. The oil that leaks into natural environments is not a single type of oil, but a mixture of various oil products, and the oil film thickness on the sea surface is uneven under the influence of wind and waves. Increasing the mixed oil film thickness dimension and the mix proportion dimension has been proposed to weaken the effect of the detection environment on the fluorescence measurement results. To preserve the relationships between the data of oil films with different thicknesses and the relationships between the data of oil films with different mixing proportions, the three-dimensional fluorescence spectral data of mixed oil films on a seawater surface were measured in the laboratory, producing a thickness-fluorescence matrix and a proportion-fluorescence matrix. The nonlinear variation of the fluorescence spectra was investigated according to the fluorescence lidar equation. This work pre-processes the data by sum normalization and two-dimensional principal component analysis (2DPCA) and uses the dimensionality reduction results as two feature-point views. Then, semi-supervised classification of collaborative training (co-training) with K-nearest neighbors (KNN) and a decision tree (DT) is used to identify the samples. The results show that the average overall accuracy of this coupling model can reach 100%, which is 20.49% higher than that of the thickness-only view. Using unlabeled data can reduce the cost of data acquisition, improve the classification accuracy and generalization ability, and provide theoretical significance and application prospects for discrimination of spectrally similar oil species in natural marine environments.