MGPNet: Rethinking multi-scale features and global attention with pre-trained model for SAR oil spill detection

Marine oil spill incidents have caused significant damage to the oceanic ecological environment, prompting the development of numerous oil spill detection methods based on SAR imagery. However, critical challenges remain, including the limitation of available datasets, confusion with look-alike substances, and variability size in target appearances, all of which significantly hinder detection accuracy. Hence, we propose a novel network, MGPNet, which rethinks multi-scale features and global attention to improve performance. MGPNet consists of five core modules: pre-trained model, encoder, bottleneck layer, decoder, and MFC. Specifically, we first leverage pre-trained models to optimize the encoder process, effectively mitigating the issue of limited training data. Second, we designed the MFSF block to capture multi-scale features and utilized the attention to suppress redundant and irrelevant information. Third, the GMFSF block is proposed to enhance feature representation and address the challenge of look-alike interference. Finally, we designed a novel multi-scale feature connection block to replace the traditional skip connections, which allows the decoder to utilize features from multiple scales and aims to enhance the ability of MGPNet to recover details during decoding. We validate the proposed MGPNet on three publicly available SAR oil spill detection datasets. Experimental results demonstrate that our method achieves state-of-the-art performance across all datasets, with mIoU gains of 0.59% on PALSAR, 0.93% on Sentinel, and 4.24% on M4D. These results highlight the potential of MGPNet for robust and accurate oil spill detection.