Abstract
The use of near-infrared (NIR) spectroscopy for material classification in recycling facilities has become increasingly prominent due to its rapid, accurate and non-destructive capabilities. However, the high cost of commercial NIR spectrometers – driven largely by their wide spectral range and high resolution – limits widespread adoption, particularly in the smaller-scale decentralized waste segregation stations. This study investigates the feasibility of reduced-range and lower-resolution NIR spectra data for accurate waste classification. To support this investigation, we constructed a dataset by collecting visible-NIR spectra (500-2500 nm) from 11 classes of real household waste materials using a laboratory-grade spectrometer with a 5nm wavelength resolution. Using this dataset, we simulate spectral degradation by systematically reducing the wavelength range and resolution. Hybrid convolutional neural networks (CNN) and support vector machines (SVM) model is trained on both the full and reduced-spec spectra to assess the impact on classification performance. The results indicate that a similar classification accuracy as the full spectral range (98.2%) can be achieved using a narrow spectral range of just 150 nm (with starting wavelengths falling within 1400–1600 nm) at a resolution of 5 nm. Additionally, the analysis shows that targeting two disjoint spectral ranges—each under 200 nm and starting at 1450 nm and 1625 nm—generally yields higher accuracy than using a single continuous range starting at 1600 nm. Overall, we present a promising direction for the development of low-cost NIR sensing systems to support broader adoption.

This publication has 14 references indexed in Scilit: