Oil spill detection by combining deep learning methods - دانشکده فنی و مهندسی
Oil spill detection by combining deep learning methods
نوع: Type: Thesis
مقطع: Segment: masters
عنوان: Title: Oil spill detection by combining deep learning methods
ارائه دهنده: Provider: Fatemeh Abbasi
اساتید راهنما: Supervisors: Dr. Morteza Heidari Mozaffar
اساتید مشاور: Advisory Professors:
اساتید ممتحن یا داور: Examining professors or referees: Dr. Hossein Torabzadeh Khorasani and Dr. Hassan Khotanlou
زمان و تاریخ ارائه: Time and date of presentation: 2026
مکان ارائه: Place of presentation: 43
چکیده: Abstract: Oil spill pollution is one of the most important environmental threats to marine ecosystems. This study aims to identify and segment oil spills in Sentinel-1 radar images using deep learning methods. For this purpose, a dataset of 1002 SAR images and their corresponding masks was used. The images were labeled into five classes including sea, land, ship, similar areas, and oil spill. In order to improve the performance of the models in the face of unbalanced data, the weighted loss function (Weighted Categorical Cross-Entropy) and the data augmentation technique (Data Augmentation) were used. In this study, five deep learning architectures including U-Net, ResNet, DeepLabV3, PSPNet, and SegNet were implemented and compared. The models were optimized using the CBAM attention module to increase focus on key areas of the image. Performance evaluation was performed based on standard metrics such as Accuracy, Dice, IoU, Precision, Recall, and F1-Score. The results showed that the PSPNet model with ResNet50 backbone provided the highest performance in all metrics and had the best balance between accuracy and sensitivity with Dice (0.78) and IoU (0.64). The SegNet model is also considered a suitable option for lighter applications with high accuracy and lower memory consumption. Overall, the findings show that combining SAR data with advanced deep learning models can provide an accurate and cost-effective tool for automatic monitoring and detection of oil spills in marine environments. By combining the CBAM attention mechanism with deep learning networks, appropriate accuracies can be achieved. By carefully studying and examining the structure of these networks, an optimal and suitable model can be created for oil spill detection and identification. Due to the rapid training of these networks, they can be used to a greater extent to identify oil spills.