LTF-MSPCNet: A synergistic approach combining attention mechanisms and local texture features for oil spill...

Marine oil spills represent a significant environmental pollution incident that severely disrupts the balance and stability of marine ecosystems. SAR (Synthetic Aperture Radar) demonstrates significant potential in oil spill detection due to its all-day and all-weather monitoring capabilities. However, existing deep learning models often overlook subtle textural features during detection, leading to frequent omissions of small-scale oil spill areas in the results. To address the existing challenges, this paper proposes LTF-MSPCNet (Local Texture Features-Multiscale Parallel Convolution). The model is built upon the TransUNet framework, integrating the LBP (Local Binary Pattern) texture feature extraction method into a multi-scale large-kernel convolution module. It further incorporates learnable feature extraction and feature reconstruction mechanisms, guiding the learned texture feature distribution to better approximate the texture characteristics of real SAR images. Furthermore, the Squeeze-and-Excitation (SE) attention mechanism is integrated after the decoder's upsampling stage to suppress background noise, thereby enhancing both the robustness and segmentation performance of the model. Experiments were conducted using carefully selected and preprocessed Sentinel-1 SAR images, including 1655 training samples and 370 testing samples. The proposed model achieves 86.46% Mean Dice and 92.13% Mean IoU, surpassing the baseline by 2.35% and 1.41%, respectively. Moreover, it also demonstrates consistent performance in practical oil spill detection, with IoU and Dice scores reaching 96.83% and 98.02%. In conclusion, the proposed model achieves accurate segmentation of oil spills and offers a novel technical pathway, providing a fresh research perspective for advancements in the field of oil spill monitoring.