Investigation on Hyperspectral Augmentation to Construction Materials Classification

Zheng Liu, Patrick Hunhold, Ziran He, Galina Polte, Elske Linß, Janice Kielbassa, Maik Rosenberger, Gunther Notni

Abstract

The recycling of construction waste faces challenges in material identification due to class imbalance in hyperspectral datasets. To address this, we propose integrating a data augmentation module into the classification workflow for construction materials using short-wavelength infrared (SWIR) reflectance spectra. Experiments were conducted with Random Forest (RF) and 1D-CNN classifiers across multi-class and binary classification tasks, where the latter targeted classes commonly confused with minority categories. Various augmentation methods were tested, with the self-attention-based WGAN (SA-WGAN) showing the most notable improvement. It increased the recall of the minority class by up to 60 and 48 percentage points in the multi-class and binary classification tasks, respectively, while maintaining stable performance on the majority classes.

Download

IMEKO-TC2-2025-008.pdf

DOI

10.21014/tc2-2025.008

IMEKO TC

TC2 - Photonics

Back to the proceedings