Rapid characterization of solid waste using near-infrared hyperspectral imaging (HSI) coupled with machine learning models has been increasingly investigated to replace the traditional time- and labor-intensive methods. However, contamination by waste-derived leachates or other fractions etc., can cause the spectra evolutions and significantly influences the identification performance, which has not been investigated before. The first attempt was made by using hyperspectral unmixing (HU) to extract the endmember components and demonstrate their contributions (abundance) to solid waste, taking the non-linear reflectance changes due to the O-H vibration of water and unclear variation associated with oil and leachates as an example. The HSI spectra of various solid waste components influenced by pure water, oil and three kinds of leachates were acquired. A novel method based on HU models, including multivariate curve resolution with alternating least squares and state-of-the-art autoencoder architectures (deep learning models), was developed to estimate the spectra of endmembers as well as their abundances in individual pixel. Their spatial distribution overview in solid waste was then yielded. The selected models were validated via an independent test data set, with lower spectral angle distance, 12.3° ± 6.5°, indicating the similarity of the predicted endmembers with real components. And the lowest root of mean square error on endmember distribution maps was 0.17. The non-linear liquid's effects by water and oil on spectra variations of solid waste were clearly illuminated. Additionally, the proposed method can extract information from mixed spectroscopic images and generate reconstructed spectra.
Keywords: Autoencoder; Endmember abundances; Endmember components; Hyperspectral unmixing; Non-linear effects; Spatial distribution.
Copyright © 2024 Elsevier Ltd. All rights reserved.