Material Identification in Complex Environments: Neural Network Approaches to Hyperspectral Image Analysis

Oct 1, 2023·

Jason Brown

Bohan Chen

Harris Hardiman-Mostow

Adrien Weihs

Andrea L. Bertozzi

Jocelyn Chanussot

· 1 min read

Cite DOI IEEE Xplore PDF DOI

Abstract

Hyperspectral imagery is often used in chemometric studies for quality sorting and recycling due to its ability to produce rich spectroscopy data. In this paper, we study plastics detection in a complex environment. In particular, we analyze hyperspectral images of three scenes with spectra in the near-infrared and visible wavelength ranges; our task is to detect plastic within the scenes. The images contain materials with high intraclass variability and significant mixing. Our novel contribution compares various methods for hyperspectral pixel classification in these complicated, real-world environments, specifically deep methods such as contrastive learning and autoencoders, as well as comparing the viability of the hyperspectral cameras’ light spectrum for the application of plastic detection.

Type

Conference paper

Publication

2023 13th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS)

This study examines plastic identification in complex hyperspectral scenes, where illumination, background materials, spectral mixing, and within-class variation make pixel-level classification substantially harder than controlled laboratory sorting. The data include both visible and near-infrared measurements, making it possible to compare the practical value of different wavelength ranges.

The paper evaluates neural feature-learning strategies—including autoencoders and contrastive learning—alongside established hyperspectral classification approaches. The comparison focuses on how well the learned representations separate plastic from confounding materials under realistic scene variability, and on which sensor spectrum is most informative for this recycling application.

Last updated on Oct 1, 2023