Unsupervised Marine Radar Oil Spill Detection via an Autoencoder–IAEA Synergistic Framework

Marine oil spill accidents pose serious threats to marine ecosystems and coastal economic development. X-band marine radar provides an effective means for nearshore oil spill monitoring, yet accurate oil film segmentation remains challenging because of grayscale overlap between oil films and sea clutter, range-dependent echo attenuation, and noise interference. To address these challenges, an unsupervised oil spill detection method integrating HBA-AE ROI extraction with IAEA fine segmentation is proposed. In the coarse detection stage, an autoencoder (AE) maps pixel grayscale intensity, local mean, and local standard deviation into a compact two-dimensional latent space. The AE-derived latent representation is further optimized by HBA to enhance oil film–sea clutter separability, after which K-Means clustering identifies candidate oil spill ROIs. In the subsequent fine segmentation stage, the extracted ROI constrains the search domain, while the Improved Alpha Evolution Algorithm (IAEA) maximizes the three-class Otsu between-class variance to determine the optimal dual thresholds for oil film delineation. Experiments using X-band marine radar data collected during the Dalian 7·16 oil spill accident were conducted against seven representative methods. The proposed framework achieved a Precision of 97.20%, an F1-score of 86.60%, and an IoU of 76.40%, demonstrating superior oil film segmentation accuracy and effective false-alarm suppression under complex sea-clutter conditions. It also maintained a favorable balance between segmentation accuracy and computational efficiency, demonstrating its potential for marine oil spill monitoring under the investigated radar observation conditions.