Data-driven fingerprinting plastic waste material using low-cost spectroscopy with data fusion

Résumé fourni par la source

Accurate and cost-effective identification of post-consumer plastic waste is essential for enabling scalable source separation for plastic waste management. While the accurate detection of plastics can be achieved using costly near-infrared (NIR) spectroscopic devices, such an approach faces diminishing effectiveness as the device’s cost and size are reduced. In this study, our objective is to explore the viability of using low-cost, narrow-band NIR spectroscopy for classifying household plastic waste using machine learning techniques. To this end, we study a handheld NIR device with the wavelength range of 740–1070 nm – which represents the third overtone in the NIR spectrum – in combination with observed categorical features to understand the performance of a low-cost data fusion approach to waste plastic identification. For analysis, we collect a dataset of spectral and categorical features for seven plastic types from household post-consumer plastic waste using the low-cost NIR spectrometer and evaluate six common classification models. In addition to conventional evaluation metrics, we propose a novel cost-sensitive metric to account for differences in misclassification costs in real-world deployment settings and illustrate its use using a simple misclassification cost matrix. Among all models, the one-versus-rest support vector machine achieved the highest overall performance (87% accuracy), demonstrating that high quality classification is feasible even with a consumer-grade, low-cost spectroscopic device. An ablation study highlighted the critical role of preprocessing and the data fusion of categorical features (29% and 9% decreases after ablation, respectively). Compared to prior work utilizing broader spectral bands, our approach yields competitive performance in a recycling context with significantly reduced hardware cost (with focus only on the third overtone region in the NIR spectrum). Given these results, our work concludes that low-cost, multi-sensor solutions that support robust classification and scalable deployment in resource-limited environments warrant further exploration.