基于多任务卷积神经网络的浮游藻类群落识别方法

Identification Method of Planktonic Algae Community Based on Multi-Task Convolutional Neural Network

针对混合浮游藻类群落离散三维荧光光谱特征识别,对比分析了简单卷积神经网络(PlainCNN)和文本卷积神经网络(TextCNN)模型对5种常见门类藻(铜绿微囊藻、斜生栅藻、菱形藻、楯形多甲藻和隐藻)混合数据的种类识别准确率及浓度测量精度。结果表明,在藻类独立识别及浓度回归分析中,PlainCNN模型对测试集的平均识别准确率和浓度输出结果的平均均方误差分别为90%和0.052,均优于TextCNN模型。为了同时实现混合藻类种类识别和浓度分析,基于PlainCNN模型提出了多任务卷积神经网络PlainCNN-MT模型。该模型对混合藻类种类识别的平均准确率提高至95%,浓度输出结果的平均均方误差降低至0.039,表明多任务卷积神经网络在浮游藻类群落识别与定量分析中更具优势。

Aiming at the identification of the characteristics of the discrete three-dimensional fluorescence spectra for the planktonic mixed algae community, the spiecies identification accuracy and concentration measurement accuracy of mixed data of five common phylum species of algae(Microcystis aeruginosa, Scenedesmus obliquus, Nitzschia sp., Peridinium umbonatum var.inaequale and Cryptomonas obovata.) are compared and analyzed by the plain convolutional neural network(PlainCNN) model and the text convolutional neural network(TextCNN) model. The results show that in the algae independent identification and concentration regression analysis, the average identification accuracy of the test set and the average mean square error of the results of the concentration output of the PlainCNN model are 90% and 0.052 respectively, which are better than that of TextCNN model. In order to realize species identification and concentration analysis of mixed algae at the same time, a multi-task convolutional neural network, i.e., PlainCNN-MT model, is proposed based on the PlainCNN model. The average accuracy of the model for the species identification of mixed algae is increased to 95%, and the average mean square error of the results of the concentration output is reduced to 0.039, indicating that the multi-task convolutional neural network has more advantages in the identification and quantitative analysis of planktonic algae community.