A lightweight spatial and spectral CNN model for classifying floating marine plastic debris using hyperspectral images

Marine plastic debris poses a significant environmental threat. In order to study and combat this pollution, efficient and automated detection methods are essential. Hyperspectral imaging and deep learning provide a robust framework for classifying floating marine plastic debris. However, deep learning approaches often suffer from high computational complexity and limited interpretability. In addition, hyperspectral images are high-dimensional data that must be analyzed efficiently. To overcome these limitations, this paper proposes the Lightweight Spatial and Spectral Hyperspectral CNN (LSS-HCNN), a deep learning model designed to enhance classification accuracy while improving computational efficiency and interpretability. The proposed model first applies spatialwise convolutions to extract spatial features from individual spectral bands, then uses spectralwise convolutions to extract relationships between spectral bands. Additionally, a Squeeze-and-Excitation (SE) block improves interpretability by focusing on the most informative spectral bands. Experiments were conducted on three hyperspectral datasets containing various materials and four dedicated floating plastic datasets, including a new plastic waste dataset. It provides images in three spectral configurations: visible-near-infrared (VIS-NIR), near-infrared-shortwave-infrared (NIR-SWIR), and a fused domain. Results show that LSS-HCNN outperforms traditional handcrafted descriptors and deep-learning models, particularly for floating marine plastics. It achieves a mean classification accuracy of 97.64% while reducing model complexity. Compared to standard 2D-CNN, it reduces the number of parameters by over 80% and Floating Point Operations (FLOPS) by a factor of 7. Moreover, SE block analysis reveals that NIR-SWIR bands contribute the most to plastic classification. This highlights LSS-HCNN as an efficient marine plastic debris classification model, supporting environmental monitoring efforts.