Jianshun Guo, Guanhua Xu, Ziyi Zhang
Port oil spill detection (OSD) is critical for maritime ecological protection and operational security. However, detection performance is frequently compromised by adverse weather and lighting conditions in complex port environments. These factors, coupled with the inherent irregular morphology and fuzzy boundaries of oil slicks, significantly confound the differentiation between oil, water, and shoreline structures. Therefore, this study develops a deep semantic driven gated refinement network named SGGR-Net to resolve the feature diffusion challenges of fragmented oil spill edges in high altitude unmanned aerial vehicle (UAV) imaging. The core innovation of this architecture lies in the sequential collaboration of the cascade semantic driven gated refinement (SGR) unit and the dual driven gated unit (DDGU). Specifically, the SGR unit establishes top down macro semantic guidance to effectively eliminate global clutter interference and spatial ambiguity. Building upon this semantically purified foundation, the subsequent DDGU employs parallel spatial and channel gating tracks combined with a depthwise separable mechanism to achieve adaptive feature harmonization and micro level fine grained boundary refinement. Furthermore, a multi scale semantic attention architecture establishes robust global contextual dependencies to adapt to diverse oil spill morphologies. To address pixel imbalance, a joint loss function combining OHEM cross entropy and log cosh dice loss is utilized to optimize segmentation accuracy. Experimental results on the public port oil spill dataset demonstrate that SGGR-Net achieves an mIoU of 93.27% and an F1-score of 96.49%. This indicates that the model exhibits superior performance in identifying ambiguous oil water interfaces and suppressing environmental interference, holding significant implications for the construction of port environmental protection systems and the maintenance of marine ecological security.