Harnessing hyperspectral imaging and deep learning for advanced e-waste classification using three spectral bands

With millions of tons of electronic devices discarded annually, accurately identifying and classifying electronic waste (e-waste) has become a critical environmental and technological challenge. This study presents a deep learning framework that leverages hyperspectral imaging (400–1000 nm) to address the limitations of traditional RGB-based sorting methods, which often fail to distinguish visually similar non-ferrous metals. We propose a modified U-Net architecture that integrates group normalization, PReLU activation, and band-wise spectral attention in skip connections to enhance spectral-spatial feature fusion. Evaluated on the Tecnalia WEEE dataset, our model achieved a classification accuracy of 92 %, a precision of 0.54, a recall of 0.56, an F1 score of 0.51, and an Intersection over Union (IoU) of 0.39. These results outperform baseline models, including the standard U-Net (accuracy: 90.15 %, IoU: 0.357), 2D Encoder-Decoder (accuracy: 85.80 %, IoU: 0.145), and 1D Encoder-Decoder (accuracy: 70.17 %, IoU: 0.018). The proposed system delivers a 23 % improvement in classification accuracy compared to RGB-based approaches, demonstrating its robustness and scalability for real-world applications. By advancing material discrimination and segmentation capabilities, this work provides a meaningful step toward fully automated and sustainable e-waste management. These advancements not only address a critical gap in recycling technology but also contribute to more sustainable waste management practices globally. By enhancing recycling efficiency, this work helps mitigate the environmental impact of growing electronic waste.