Classification of soil-embedded colored plastics of many types with NIR hyperspectral high-resolution imaging system and multi-scale 3D deep learning

The accumulation of various colored plastics waste in terrestrial ecosystems poses an escalating threat to environmental sustainability, demanding effective and precise classification methods. However, current machine vision techniques often fail due to significant interference caused by plastic colorants. Their performance in real-world scenarios is further compromised by complex soil backgrounds, including soil heterogeneity and varying moisture content. To overcome these limitations, we present a novel NIR hyperspectral imaging system, integrated with a specially designed MS3D-Net, for high-precision classification of soil-embedded plastics. The MS3D-Net model simultaneously performs pixel-level segmentation and polymer classification for the identification of 11 plastic types within complex soil backgrounds. We evaluated the system's robustness across three distinct heterogeneous soil types (loam, clay, and sand) under varying moisture gradients, demonstrating stable performance against matrix interference. To validate applicability in realistic field conditions, the system was tested on plastics subjected to aging, oxidation, fragmentation, and biofouling embedded in natural soils containing complex impurities such as plant roots and residues, achieving an average precision of 0.89. The method successfully recognized irregular consumer plastic products collected from the environment. Our approach provides a robust solution for identifying diverse plastic polymers in complex soil environments, offering technical support for environmental sustainability and plastic waste management. • A custom-built NIR-HSI system obtains high-resolution hyperspectral images. • Multi-scale 3D deep network accurately classifies soil-embedded plastics. • 0.89 precision across 11 plastic types under aged, cracked, biofouled conditions.