Classification of marine plastic debris using hyperspectral imaging and band selection: A patch-based and pixel-based fusion approach

Abstract

Marine plastic pollution poses significant ecological, economic, and social challenges, requiring innovative monitoring and identification solutions to support effective mitigation and management strategies. Hyper-spectral imaging and artificial intelligence have proven to be valuable tools in detecting and identifying macroplastics in aquatic environments. Despite numerous studies focusing on deep learning approaches, many existing models remain computationally heavy and lack adaptability for real-world on-board processing on energy-constrained platforms like drones. This drawback limits their applicability for large-scale monitoring and requires models that are both precise in their predictions and lightweight for efficient computation. First, to improve plastic-type classification performance, this paper proposes an uncertainty-aware fusion approach where the recently proposed patch-based Lightweight Spatial and Spectral Hyperspectral Convolutional Neural Network (LSS-HCNN) is fused with a pixel-based Random Forest (RF) classifier. Second, to improve computational efficiency, this paper investigates two band selection methodologies based on LSS-HCNN Squeeze-and-Excitation (SE) block weights and RF feature importances respectively. This study evaluates the classification of five common polymers (HDPE, LDPE, PET, PP, and PS) supplemented by natural organic matter and background materials. To address material heterogeneity at object boundaries, we evaluate the approach on both pure and mixed-material regions. The results show that LSS-HCNN consistently outperforms traditional Machine Learning (ML) methods, improving performance by more than 4% over the accuracy provided by RF. The proposed uncertainty-aware fusion successfully enhances classification accuracy, achieving 97% on hyperspectral images of plastic debris. Furthermore, a subset of six selected bands, identified by the elbow method as the optimal accuracy-efficiency trade-off, maintains 90% accuracy while reducing computational demands by more than 20 times fewer parameters and floating-point operations. Our findings provide a pathway towards lightweight, accurate, and adaptable models for real-time plastic debris monitoring in aquatic environments.