A knowledge-based, validated classifier for the identification of aliphatic and aromatic plastics by WorldView-3 satellite data

Abstract

Although the CH chains of petroleum derivatives display unique absorption features in the short-wave infrared (SWIR), it is a challenge to identify plastics on terrestrial surfaces. The diverse reflectance spectra caused by chemically varying polymer types and their different kinds of brightness and transparencies, which are, moreover, influenced further by the respective surface backgrounds. This paper investigates the capability of WorldView-3 (WV-3) satellite data, characterized by a high spatial resolution and equipped with eight distinct and relatively narrow SWIR bands suitable for global monitoring of different types of plastic materials. To meet the objective, hyperspectral measurements and simulations were conducted in the laboratory and by aircraft campaigns, based on the JPL-ECOSTRESS, USGS, and inhouse hyperspectral libraries, all of which are convolved to the spectral response functions of the WV-3 system. Experiments further supported the analyses wherein different plastic materials were placed on different backgrounds, and scaled percentages of plastics per pixel were modeled to determine the minimum detectable fractions. To determine the detectability of plastics with various chemical and physical properties and different fractions against diverse backgrounds, a knowledge-based classifier was developed, the routines of which are based on diagnostic spectral features in the SWIR range. The classifier shows outstanding results on various background scenarios for lab experimental imagery as well as for airborne data and it is further able to mask non-plastic materials. Three clusters of plastic materials can clearly be identified, based on spectra and imagery: The first cluster identifies aliphatic compounds, comprising polyethylene (PE), polyvinylchloride (PVC), ethylene vinyl acetate copolymer (EVAC), polypropylene (PP), polyoxymethylene (POM), polymethyl methacrylate (PMMA), and polyamide (PA). The second and third clusters are diagnostic for aromatic hydrocarbons, including polyethylene terephthalate (PET), polystyrene (PS), polycarbonate (PC), and styrene-acrylonitrile (SAN), respectively separated from polybutylene adipate terephthalate (PBAT), acrylonitrile butadiene styrene (ABS), and polyurethane (PU). The robustness of the classifier is examined on the basis of simulated spectra derived from our HySimCaR model, which has been developed in-house. The model simulates radiation transfer by using virtual 3D scenarios and ray tracing, hence, enables the analysis of the influence of various factors, such as material brightness, transparency, and fractional coverage as well as different background materials. We validated our results by laboratory and simulated datasets and by tests using airborne data recorded at four distinct sites with different surface characteristics. The results of the classifier were further compared to results produced by another signature-based method, the spectral angle mapper (SAM) and a commonly used technique, the maximum likelihood estimation (MLE). Finally, we applied and successfully tested the classifier on WV-3 imagery of sites known for a high abundance of plastics in Almeria (Spain), Cairo (Egypt), and Accra, (Ghana, West Africa). Both airborne and WV-3 data were atmospherically corrected and transferred to “at-surface reflectances”. The results prove the combination of WV-3 data and the newly designed classifier to be an efficient and reliable approach to globally monitor and identify three clusters of plastic materials at various fractions on different backgrounds.