Publication

Abstract : The proliferation of plastic debris in terrestrial and aquatic environments poses significant ecological and monitoring challenges worldwide. Hyperspectral imaging (HSI) provides fine-grained spectral signatures that enable the reliable identification of plastic types; however, its high dimensionality introduces redundancy, computational burden, and susceptibility to the Hughes phenomenon. To address these issues, this study proposes a two-phase framework that integrates Superb Fairy-Wren Optimization (SFWO) for band selection with a band-specific pooling spectral Convolutional Neural Network (BSPS-CNN) for classification. In the first phase, the SFWO operates as a wrapper-based metaheuristic that selects an informative and non-redundant band subset by optimizing a fitness function that combines the signal-to-noise ratio, variance, and inter-band correlation. In the second phase, the BSPS-CNN employs multi-scale one-dimensional convolutions, concatenation, and band-specific global average pooling to extract discriminative spectral features from the reduced hyperspectral cube. The resulting framework improves the plastic debris classification accuracy while substantially reducing the input dimensionality and computational cost. Experiments on hyperspectral plastic debris datasets demonstrated that the proposed approach outperformed existing band selection and deep learning methods in terms of accuracy, efficiency, and robustness.