Deep Learning-Based Hyperparameter Optimization for Enhanced Segmentation of RGB Images in Oil Spill Detection Within Port Environments

Abstract

Accurate detection of oil spills in dynamic port environments remains critical for timely response and effective environmental protection. Recent studies have explored the application of deep learning models to high-resolution RGB imagery; however, challenges persist due to the complex visual characteristics of ports, including dynamic water surfaces, infrastructural occlusions, and heterogeneous backgrounds. Traditional detection approaches have demonstrated limited robustness under such conditions, often resulting in elevated rates of false positives and missed detections. In this study, a semantic segmentation model based on the DeepLabV3 architecture was developed and optimized for oil spill detection using drone-acquired RGB imagery. Model performance was enhanced through the integration of convolutional neural networks, transfer learning, and systematic hyperparameter tuning involving batch size, learning rate, and training epochs. Evaluation was conducted through five-fold cross-validation to ensure generalization capability, and the results were benchmarked against established segmentation models. The optimized model achieved a validation accuracy of 96.53%, with an F1-score of 0.9557 and an Intersection over Union (IoU) of 0.8643. Cross-validation yielded a mean accuracy of 89.35% (±0.0103), precision of 88.52% (±0.0081), recall of 85.26% (±0.0138), F1-score of 86.45% (±0.0122), and IoU of 76.97% (±0.0177), indicating consistent performance across all partitions. Per-class performance further supported model reliability, with F1-scores of 0.92 for oil, 0.91 for other objects, and 0.97 for water, and corresponding IoU values of 0.8569, 0.8273, and 0.9469, respectively. When compared to baseline models such as Fully Convolutional Networks (FCN, IoU: 83.48%), U-Net (IoU: 78.80%), and MFSCNet (IoU: 86.24%), the proposed DeepLabV3-based architecture demonstrated superior segmentation accuracy. These findings indicate the model's potential for real-time deployment in high-risk maritime zones, offering significant contributions to environmental monitoring, rapid incident response, and operational decision-making.