Diverse Region-Based ViT Model for Marine Oil Spill Detection Based on Satellite Multispectral Data

Marine oil spill incidents pose significant threats to both the marine ecosystem and coastal economies. Timely and precise detection of oil spills is essential for efficient response and management of such events. This study introduces a Diverse Region-Based Vision Transformer (DR-ViT) model, designed to overcome the limitations of conventional approaches in detecting variable oil spill patterns within complex marine environments. The model achieves accurate identification of oil spill pixels by incorporating multi-directional regional inputs and integrating spectral and spatial features from each region. Evaluated using multispectral data from the Haiyang-1C/D (HY-1C/D) satellites across the Yellow Sea, Persian Gulf, and South China Sea, the model achieves an F1-Score of 0.880 in standard conditions, surpassing traditional machine learning (e.g., SVM) and deep learning baselines (e.g., CNN, ViT). In challenging scenarios involving floating algae and cloud shadows, DR-ViT maintains strong performance with F1-Scores of 0.798 and 0.814, respectively, demonstrating resilience and cross-scene adaptability. The multi-directional fusion approach effectively resolves the shortcomings of fixed-region methods, enhancing detection of irregular oil slicks across diverse marine environments. While computationally intensive, DR-ViT offers a scalable framework for real-time monitoring, with potential for optimization. This work advances remote sensing technology, providing a powerful tool for marine environmental protection and emergency management.