Using hyperspectral imaging to identify optimal narrowband filter parameters for construction and demolition waste classification

TL;DR

This study demonstrates that using two carefully selected narrowband filters in the near-infrared region with standard cameras can effectively classify construction and demolition waste, reducing computational complexity while maintaining high accuracy.

Contribution

It introduces a method for selecting optimal narrowband filter parameters for waste classification using hyperspectral data and a simple classifier, enabling practical real-time sorting.

Findings

Two wavelengths beyond RGB suffice for high-accuracy classification

Optimal filter wavelengths are approximately 650-750 nm and 850-1000 nm

Near-infrared regions are crucial for material discrimination

Abstract

Hyperspectral imaging (HSI) is widely applied in various industries, enabling detailed analysis of material properties or composition through their spectral signatures. However, for classification of construction and demolition waste (CDW) materials, HSI is impractical since real-time sorting requires rapid data acquisition and lightweight classification. Instead, fitting selected narrowband filters onto standard cameras can achieve comparable results with substantially reduced computational overhead. In this study, reflectance data of common CDW materials were recorded using a hyperspectral camera, and a multilayer perceptron classifier was employed to evaluate different feature sets. The findings indicate that adding only two wavelengths beyond the RGB channels is sufficient for high-accuracy classification, with optimal filter central wavelengths identified at approximately 650-750 nm and 850-1000 nm across the tested bandwidths (5-50 nm) highlighting the importance of near-infrared regions for material discrimination.