Indexed in

License and use

Citations

Cited 25 times in

Altmetrics

Grant support

This research was supported by the China Scholarship Council, (Grant number 201906220224) , the Sentinel4marine plastic waste project (Grant number: 50EE1269) and the EnMAP project (Grant number: 50EE1923) funded by the German Federal Ministry for Eco-nomic Affairs and Energy (BMWi) . The airborne HyMap recordings were conducted by the German Aerospace Center (DLR) . The HySpex Mjolnir data recordings were conducted and geometrically corrected by Chris-tian Mielke, Friederike Kastner, and Nicole Koellner of the GFZ. Shanyu Zhou wishes to thank especially Rainco Xiao for his sophisticated sup-port. Finally, the authors would like to thank the anonymous reviewers for their efforts and constructive suggestions to improve the clarity, precision, and relevance of this article.

Analysis of institutional authors

Share

Identifying distinct plastics in hyperspectral experimental lab-, aircraft-, and satellite data using machine/deep learning methods trained with synthetically mixed spectral data

Publicated to:Remote Sensing Of Environment. 281 113263- - 2022-09-12 281(), DOI: 10.1016/j.rse.2022.113263

Authors: Zhou, Shanyu; Kaufmann, Hermann; Bohn, Niklas; Bochow, Mathias; Kuester, Theres; Segl, Karl

Affiliations

Abstract

The growing production and use of plastics are becoming a serious progressive issue and people pay increasing attention to the effects of plastics on ecosystems and human health. The availability of hyperspectral data from space sensors inspired us to study the feasibility to detect and identify different types of plastics in aircraft-, Goafen-5 (GF-5) and PRISMA satellite data by means of deep-, and machine learning models trained with spectral signatures. In this context, various inhouse and public spectral libraries are used to create a compre-hensive database with mixed pixels of different plastic and non-plastic materials. The endmembers of plastic types involved in this study are polyethylene (PE), polypropylene (PP), polyvinyl chloride (PVC), polyethylene terephthalate (PET) and polystyrene (PS), covering 95% of the global production. Additionally, some important varieties of industrial plastics types such as acrylonitrile butadiene styrene (ABS), ethylene vinyl acetate (EVA), polyamide (PA), polycarbonate (PC), and polymethyl methacrylate (PMMA) were included in the investigations. Different samples with varying optical properties (color, brightness, transmissivity) have been selected for each plastic type. As non-plastic materials we have chosen spectra of vegetation, rocks, soils and minerals contained in the public US libraries (ECOSTRESS and USGS). The number of spectra for the training of the deep learning and machine learning models was enlarged by a random linear mixing method and the resulting database was separated into a training and a test group for subsequent multi-label classification. Algorithms selected are a convolutional neural network (CNN), random forest (RF) and support vector machine (SVM). To investigate the transferability to any hyperspectral image data obtained by air-, and spacecraft sensors, we opted for a unifi-cation of the spectral response functions (SRF) and the spectral sampling intervals of all data. Validation is accomplished based on the test group of the spectral database, and tested by controlled laboratory and aircraft experiments recorded over surfaces with varying background materials. Results are further analyzed for the influence of different noise quantities and abundance levels. The performance of the three models is roughly balanced for the validation of the spectral data with an overall accuracy of 97%, 96%, and 95% for the CNN, RF, and SVM, models respectively. In the controlled lab experiments, various accuracy indicators, such as the recall rates and the comprehensive metrics F1-score, OA, and Kappa suggest the RF classifier as the most robust one, followed by the SVM and CNN models. As for the evaluation of the aircraft data from controlled experiments, the RF further outperforms the other two models, behaving most robustly and reliably against conditions with un-known plastics and unknown background surfaces. Thus, the RF was used to classify the ten types of plastics mentioned above in one GF-5 and two PRISMA satellite recordings of the same area. In comparison of both sensor systems, the RF produced high quality and transferable results for detecting plastic mainly related to green-houses, sport fields, photovoltaic constructions and industrial sites that are discussed in detail in this paper.

Keywords