Intelligent Visible-Near Infrared Micro-Hyperspectral Sensing System for Rapid Chemical Mapping of Microplastics and Metal Oxides

Rapid, non-destructive, and accurate chemical mapping of microscopic materials is critical for advancing chemical analysis and related industries. However, conventional techniques like scanning electron microscopy with energy-dispersive X-ray spectroscopy (SEM-EDS) and Raman microscopy are often limited by low throughput and high costs. To overcome these barriers, we report the development of an intelligent sensing platform that integrates low-cost visible and near-infrared (Vis-NIR) micro-hyperspectral imaging with a custom-designed deep learning architecture. The core of our innovation is a patch-based, spatial-spectral strategy implemented through a custom-designed multi-attention 3D convolutional neural network with residual connections. This approach effectively compensates for the low chemical specificity of broad Vis-NIR spectra by learning subtle, high-dimensional joint features. The platform's power is demonstrated by its ability to classify a challenging set of eight chemical species, including spectrally indistinct microplastics (polystyrene and poly(methyl methacrylate)) and various metal oxides, with 97.35% accuracy. The high-fidelity chemical maps of complex, multi-component agglomerates were rigorously validated against SEM-EDS, confirming the model's robustness. Critically, our non-destructive optical method achieves this with a throughput several orders of magnitude higher than SEM-EDS. This work provides a powerful and versatile tool for the high-throughput characterization of diverse materials, including metal oxide catalysts, environmental contaminants like microplastics, and other complex heterogeneous systems, with broad applications across scientific and industrial domains.