High-accuracy quantitative segmentation of Sub-10 μm microplastics using hyperspectral deep learning

Xinwei Dong, Tao Zhang, Fei Guo, Yansheng Liu, Fuxin Zheng, Guoxiao Xu, Guofu Wang,
High-accuracy quantitative segmentation of Sub-10 μm microplastics using hyperspectral deep learning,
Microchemical Journal,
Volume 219,
2025,
115862,
ISSN 0026-265X,
https://doi.org/10.1016/j.microc.2025.115862.
(https://www.sciencedirect.com/science/article/pii/S0026265X25032102)
Abstract: The growing global concern over microplastics, particularly those smaller than 10 μm, has highlighted the need for reliable methods to detect and quantify these particles. Traditional imaging and chemical analysis techniques, such as scanning electron microscopy (SEM) and Raman spectroscopy, face significant limitations in high-throughput analysis and struggle to handle the complex, mixed morphology of microplastics. In response to these challenges, this study introduces a novel patch-based deep learning framework, 3DCNN-GSTM (3D Convolutional Neural Network with Global Spectral Transformer Module), integrated with Micro-Hyperspectral Imaging (Micro-HSI). Employing a patch-based approach to integrate spatial and spectral information while incorporating contextual data from surrounding pixels, this framework overcomes traditional pixel-wise classification limitations, achieving near-perfect average accuracies of 99.39 % and ensuring robust, high-fidelity segmentation in complex multi-component samples. Rigorous validation experiments affirm the model's efficacy in accurately segmenting and quantifying microplastics of approximately 5 μm in size across single-component and mixed samples. In multi-component mixtures, the model successfully differentiated morphologically similar particles, achieving strong linear correlations (R2 > 0.87) between the predicted pixel-wise class proportions and the true concentration ratios. This work represents a significant advancement in microplastic detection, providing an automated, reliable method for precise quantification of microplastic composition, spatial distribution, and abundance.
Keywords: Sub-10 μm microplastics; Image segmentation; Microscopic hyperspectral imaging; Artificial intelligence; Patch-based classification