Marine oil spill detection using synthetic aperture radar (SAR) imagery is vital for environmental monitoring and emergency response. However, accurate identification of oil spill regions remains challenging due to complex sea surface backgrounds, significant variations in spill scale and morphology, and blurred boundaries between contaminated and clean water areas. Existing methods struggle to jointly model multi-scale contextual information and fine-grained structural details, exacerbating these challenges. To address these issues, a feature-guided interconnection-based network for sea surface oil spill detection is proposed. This network enhances feature representation across multiple scales while preserving boundary integrity and offers an effective experimental case for applying artificial intelligence techniques to real-world marine engineering problems.

A feature-guided interconnection network, named FGINet, is proposed for oil spill detection in SAR images. The network employs Res2Net-50 as its backbone to extract hierarchical multi-scale features. To address scale variation and contextual ambiguity, an enhanced feature-guided aggregation (EFA) module is designed. This module uses dilated convolutions with different receptive fields to capture oil spill structures at multiple scales. In addition, an adjacent feature interconnection (AFI) module is introduced to facilitate effective interaction between adjacent semantic layers through attention mechanisms, allowing for refined boundary modeling and suppression of background interference. The proposed network is trained and evaluated on two publicly available SAR oil spill datasets from the Gulf of Mexico and the Persian Gulf. The model performance is quantitatively evaluated using the structural measure (S_α), mean absolute error ( $\mathcal{M}$), E-measure (E_ξ), and F-measure (F_β). Comparative experiments are conducted against 11 state-of-the-art methods, and comprehensive ablation studies are performed to analyze the contribution of each proposed module.

Experimental results demonstrate that FGINet consistently outperforms all other comparison methods across both datasets. On the Gulf of Mexico dataset, the proposed approach achieves an S_α value of 0.7945 and F_β score of 0.7226, while reducing the $\mathcal{M}$ to 0.0635. Similar performance gains are observed on the Persian Gulf dataset, indicating strong robustness and generalization capabilities under different marine environments. Qualitative visual comparisons reveal that FGINet produces more accurate oil spill localization, clearer oil stain boundaries, and fewer false detections, particularly in low-contrast backgrounds, narrow oil spill structures, and multiple spill regions. Furthermore, ablation experiments confirm that the EFA and AFI modules contribute indispensably to performance improvement and that removing any module leads to noticeable degradation in detection accuracy.

This study demonstrates that the proposed FGINet effectively addresses key challenges in SAR-based oil spill detection by combining multi-scale feature aggregation with feature-guided interconnection mechanisms. Experimental results confirm its superior accuracy and robustness over existing approaches. Beyond its technical contributions, the framework provides a valuable experimental teaching case, allowing students to gain hands-on experience in deep-learning-based remote sensing analysis and model optimization. The methodology shows strong potential for real-world marine environmental monitoring and emergency response applications, offering a practical reference for future research on intelligent ocean observation systems.