摘要
微塑料作为一种新型污染物,引起的污染问题在当今越发受到研究人员的关注。拉曼光谱有着无损样品,光谱特征峰位有代表性,便于识别样品种类的优点,一直以来是生物化学分析领域热门的检测方法之一。深度学习近年在特征提取、目标检测等领域的发展成就引人瞩目。针对准确高效的微塑料检测方法问题,探索了基于马尔可夫变迁场(MTF)变换的拉曼光谱和卷积神经网络的微塑料识别可行性。采集了11种微塑料样品的拉曼光谱,每种样品采集100条光谱,并通过光谱平移、添加噪声、光谱叠加等方式数据强化扩充了光谱数据集,利用MTF将一维拉曼光谱数据转换为彩色二维图像,获得二维图像形式的光谱数据集。设计一种主体为多个小尺寸卷积核的卷积神经网络(SSMKB-CNN)模型,包括1个输入层、6个卷积层、2个池化层、1个平坦层、2个全连接层与1个输出层。选择Dropout与带权重衰减的自适应矩估计优化器以避免过拟合现象,采用阶梯式学习速率保证模型充分习得数据深层特征。分类性能的评价标准采用混淆矩阵、精确率、召回率、F1分数与整体准确率。模型的分类结果与基于二维图像光谱数据集的AlexNet、VGG16和ResNet50三个深度卷积神经网络模型和基于一维光谱数据集的随机森林(RF)、K-最邻近(KNN)和三种核函数(rbf、Linear、Poly)的支持向量机(SVM)机器学习分类器的分类结果进行对比。通过分析训练曲线及混淆矩阵的分类结果,四种CNN模型的损失与准确曲线走势平稳,均能达到良好的训练效果,其中提出的SSMKB-CNN模型准确率达到了97.04%,综合精确率、召回率和F1分数分别为97.05%、95.06%和97.02%,均明显优于用于比较的另外三种经典神经网络模型以及三种机器学习分类器。一轮训练时长为9s,训练时间少于三种CNN模型,综合分类性能最佳。实验结果表明,基于结合MTF变换的拉曼光谱的SSMKB-CNN模型能够准确高效地提取光谱特征并作出种类预测,利用拉曼光谱实现微塑料样品的定性识别,可以为海水中微塑料实际样品的准确客观的识别技术提供方法新思路参考。
In recent years,seawater pollution caused by microplastic waste has caught more and more attention.Raman spectroscopy,a non-destructive detection technique,has representative spectral characteristic peaks,making it easier to identify unknown samples.It has always been one of the popular detection methods in biochemical analysis.Deep learning has made remarkable achievements in feature extraction,target detection,and other fields.The feasibility of Raman spectroscopy based on the Markov transition field(MTF)combined with a convolution neural network(CNN)was explored for the accurate and efficient detection of microplastics.The Raman spectra of eleven types of microplastic samples were collected,and 100spectra were collected for each sample;then,the spectral dataset was expanded through data augmentation.The one-dimensional Raman spectral data was converted into two-dimensional images using a Markov transition field to obtain a two-dimensional image spectral dataset.A small-sized multiple-kernel-based convolutional neural network(SSMKB-CNN)model with continuous smallscale convolutional kernels is designed,including one input layer,six convolutional layers,two pooling layers,one flattened layer,two fully-connected layers,and one output layer.The classification performance of the model is compared with the classification results of AlexNet,VGG16,and ResNet50deep convolution neural network models based on a two-dimensional MTF image spectral dataset,along withthree classical machine learning classifiers based on a one-dimensional spectral dataset,including K-nearest neighbors(KNN),random forest(RF)and support vector machine(SVM)with three kernel functions(rbf,Linear and Poly).It could be seen from the training curves and the classification confusion matrix that the loss and accuracy curves of the four CNN models are stable and can achieve a good training effect.The accuracy rate of the proposed SSMKB-CNN model reaches 97.04%,and the macroprecision rate,recall rate,and F1-score are 97.05%,95.06%,and 97.02%,respectively,which is superior to the other three CNN models used for comparison and the three machine learning classifiers.Each training epochconsumes 9seconds,less than the three CNN models.Overall,the proposed SSMKB-CNN model has the best classification performance.The experimental results show that the Raman spectrum and SSMKB-CNN model combined with MTF can accurately and efficiently extract spectral features and make precise predictions,and the qualitative identification of microplastic samples using the Raman spectrum is realized.It can provide a method reference for the recognition technology of actual microplastic contaminants in seawater.
作者
张蔚
冯巍巍
蔡宗岐
王焕卿
闫奇
王清
ZHANG Wei;FENG Wei-wei;CAI Zong-qi;WANG Huan-qing;YAN Qi;WANG Qing(Yantai Research Institute of Harbin Engineering University,Yantai 264000,China;CAS Key Laboratory of Coastal Environmental Processes and Ecological Remediation,Yantai Institute of Coastal Zone Research,Chinese Academy of Sciences,Yantai 264003,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2024年第9期2420-2427,共8页
Spectroscopy and Spectral Analysis
基金
国家重点研发计划项目(2019YFD0901101)
山东省重点研发计划项目(2022CXPT019)资助。
关键词
马尔可夫变迁场
拉曼光谱
微塑料
神经网络
Markov transition field
Raman spectrum
Microplastics
Convolutional neural network
