协同增强技术驱动的环境微塑料样本光谱识别

Spectral Identification of Environmental Microplastic Samples Driven by Collaborative Enhancement Technology

摘要: 目的：微塑料样本在光谱识别中同时受到样本多样性及样本数量分布不平衡等困扰。环境中微塑料样本不仅成分繁多，且处于不同的老化阶段，直接导致检测模型泛化能力薄弱。环境微塑料光谱的精准识别已成为该领域的核心难题。方法：本文提出一种物理–生成式协同增强模型，通过融合两类增强样本来扩充模型训练集，进而提升识别模型的泛化性能，最终达成样本的有效扩充与精准识别的双重目标。增强的样本可以通过物理特征增强方法生成，还可以由改进的Wasserstein对抗网络(CWGAN-GP-SAM)生成，该网络引入光谱角匹配(SAM)确保生成样本的可靠性。在建模过程中融合两类增强样本来构建扩充样本集，解决了训练样本分布不平衡与样本多样性问题。结果与讨论：为验证增强策略的性能，选用卷积神经网络(Convolutional Neural Network, CNN)、长短期记忆网络(Long Short-Term Memory, LSTM)、随机森林(Random Forest, RF)及XGBoost等模型进行实验对比。结果表明，四个模型在原始数据集上的性能表现具有一致性，其中CNN模型准确率为88.54%，LSTM模型准确率为87.61%；经单一增强策略优化后，LSTM模型通过生成式增强准确率提升至91.33%，较原始数据提升4.25%；而融合物理特性与CWGAN-GP-SAM的增强策略展现出最优性能，其中CNN模型准确率提升至92.26%，较原始数据提升4.20%，LSTM模型准确率达91.64%，提升幅度达4.60%。结果显示协同增强技术显著增强了环境微塑料样本的精确识别能力。

Abstract: Objective: In the spectral detection of microplastic samples, the analysis is concomitantly plagued by insufficient sample diversity and the scarcity of specific sample types. Microplastic samples in natural environments exhibit highly diverse chemical compositions and undergo distinct stages of aging, which consequently results in poor generalization performance of spectral detection models. Thus, achieving effective data augmentation and accurate identification of environmental microplastic spectra has emerged as a pivotal challenge in this research field. Methods: A physical-generative collaborative augmentation model is proposed. By fusing two types of augmented samples to expand the model training set, the generalization performance of the model is markedly enhanced, thereby ultimately attaining the dual goals of effective sample expansion and accurate spectral identification for microplastics. One subset of the augmented samples is generated via a physical feature augmentation approach, while the other is produced by an improved conditional Wasserstein generative adversarial network (CWGAN-GP-SAM), which explicitly incorporates spectral angle matching (SAM) as a constraint to guarantee the validity and reliability of the generated samples. In the model construction process, the two categories of augmented samples are integrated to construct an expanded sample set, which comprehensively alleviates the longstanding issues of training sample scarcity and insufficient sample diversity in the field of microplastic spectral detection. Results and Discussions: To assess the effectiveness of the proposed enhancement strategies, comparative experiments were performed on a suite of baseline models, including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Random Forest (RF), and XGBoost. The results indicate consistent performance across the four models on the original dataset, with the CNN and LSTM models achieving accuracies of 88.54% and 87.61%, respectively. Following optimization with a single augmentation strategy, generative augmentation improved the LSTM model's accuracy to 91.33%, representing a 4.25% increase compared to the original data. Furthermore, the augmentation strategy integrating physical characteristics with CWGAN-GP-SAM demonstrated the optimal performance. Under this strategy, the accuracy of the CNN model increased to 92.26% (a 4.20% improvement over the original data), while the LSTM model reached 91.64% (a 4.60% improvement). These findings demonstrate that the synergistic augmentation technology significantly enhances the precise identification capability of environmental microplastic samples.