Influence of Wavelength Selection and Data Preprocessing on Near-Infrared-Based Classification of Demolition Waste

Abstract

The separation of demolition waste into three fractions-wood (required purity >90%), plastic (required purity >80%), and stone (no requirement)—has been investigated. The materials are measured with diffuse near-infrared reflectance spectroscopy and classified with linear discriminant analysis (LDA). To speed up the classification, simulated annealing extracts the six most discriminating wavelength regions for each preprocessing technique. For improvement in the classification results, several preprocessing techniques are investigated. Both the reflectance R and log 10(1/R) are investigated. The influence of the so-called wavelength shift (radiation that does not pass a filter perpendicular and shifts maximally 6 nm to lower wavelength) is accounted for during wavelength selection. Preprocessing methods that remove spectral offset differences and remove differences in peak heights give the best classification improvement. Modified standard normal variate (SNV) preprocessing, applied on the reflectance R, is the best preprocessing technique. The modification is the addition of the mean spectral value after the application of standard SNV preprocessing. The mean spectral value contains some additional information that is used by the (LDA) algorithm to improve the classification performance. The influence of the so-called wavelength shift effect is minimal.