Enhancing marine oil spill detection through dynamic adaptive knowledge distillation with spectral mask superpixel

Deep learning (DL) has advanced marine oil spill detection by leveraging hyperspectral images (HSI) to capture rich contextual and semantic information. However, the scarcity of labeled samples limits DL's potential. To address this, we propose a dynamic-adaptive knowledge distillation method with spectral mask superpixels (DAKD-SMS), which automatically extracts spatial-spectral features from unlabeled data to train and refine the model. DAKD-SMS integrates four key components: 3D data transformation, a dynamic vision transformer (ViT) network, superpixel generation with spectral index masks, and scale-adaptive knowledge distillation. First, six standard oil spill spectral indices are used to create feature masks, enabling rapid superpixel segmentation of spill regions. Knowledge distillation then produces self-labeled samples using oil-specific spectral indices, optimizing the ViT network's hyperparameters and outperforming traditional self-attention mechanisms. The scale-adaptive distillation module calculates spatial-spectral joint distances (SSJD) between unlabeled superpixels and target classes, generating soft-labeled superpixel sets to estimate class probabilities. The model is trained by minimizing cross-entropy loss between predictions and soft/hard labels. Additionally, 3D transformation augments limited HSI data through predictive fusion. Extensive experiments have demonstrated that the proposed model achieves outstanding oil spill detection accuracies of 98.84%, 93.63%, and 96.67% on three datasets, outperforming current methods. DAKD-SMS addresses the bottleneck of labeled data and enhances oil spill detection through robust spatial-spectral feature extraction and adaptive learning.