RGB and RGNIR image dataset for machine learning in plastic waste detection

Area of Science:

• Environmental Science
• Computer Science
• Materials Science

Background:

• Plastic waste poses a significant environmental challenge, necessitating advanced sorting technologies.
• Spectral imaging offers potential for plastic identification but faces limitations in cost, complexity, and resolution.
• Machine learning (ML) shows promise for analyzing plastic waste data, but requires comprehensive datasets.

Purpose of the Study:

• To address the lack of publicly available datasets combining RGB and near-infrared (NIR) spectral data for plastic waste detection.
• To introduce a novel, comprehensive dataset designed for training ML models for plastic waste identification.
• To facilitate advancements in automated plastic waste sorting and management.

Main Methods:

• Collected onshore images of plastic waste using standard RGB and Red-Green-Near-Infrared (RGNIR) spectral channels.
• Captured 405 images per dataset along riverbanks and beaches.
• Pre-processed and annotated a total of 1,344 plastic waste objects for ML model training.

Main Results:

• Developed two distinct datasets: one RGB and one RGNIR spectral dataset.
• Successfully annotated 1,344 plastic waste objects, providing detailed feature information.
• The dataset uniquely integrates RGB and NIR spectral data, offering richer information than standard RGB alone.

Conclusions:

• The introduced dataset provides a valuable, unique resource for researchers in plastic waste detection.
• This dataset is expected to enhance the performance of ML models in identifying and classifying plastic waste.
• The findings encourage further research into spectral imaging and ML for improved plastic waste management strategies.