The Master's thesis of researcher Iman Ali Mohammed was discussed at the College of Engineering, University of Basra, Department of Computer Engineering, under the supervision of Professor Dr. Ghaydaa Abdul-Razzaq Suhail. The thesis is titled "Oil Spill Detection in Water Bodies Using Deep Learning and Image Analysis:
A Strategy for Environmental Protection
 " It includes...
Oil spill detection in marine, ports and aquatic environments is a critical environmental challenge due to the severe ecological and economic impacts caused by oil contamination. Therefore, developing accurate and reliable systems for detecting and analyzing oil spills using advanced deep learning techniques has become essential for effective environmental monitoring. This thesis proposes an integrated framework for oil spill detection in terms of classification, and segmentation using RGB remote sensing and drone images. Several deep learning models and hybrid machine learning techniques were investigated to improve detection accuracy, address data imbalance, and support reliable environmental monitoring systems.
    Initially, transfer learning models based on VGG16 and VGG19 were developed for oil spill image classification. The results showed that VGG16 and VGG19 were developed for oil spill image classification. The results showed that VGG16 achieved an accuracy of 95.67%, while VGG19 achieved 96.06% after 70 training epochs. Feature analysis techniques including t-SNE nonlinear dimensionality reduction technique and heatmap Grad-CAM (Gradient weighted Class Activation Mapping) were applied to visualize feature distributions and interpret model decisions. Furthermore, a lightweight advanced MobileNetV2 was enhanced using Squeeze-and-Excitation (SE) attention mechanism. The baseline MobileNetV2 achieved 94.49% accuracy, while the enhanced model achieved 95.28%. K-Fold cross-validation was used to ensure model robustness and generalization.
In addition, a new ResNet50-OSD hybrid framework combining ResNet50 with traditional machine learning algorithms was developed. The ResNet50 baseline achieved 92% accuracy, while ResNet50-SVM (Support Vector Machine) with a data augmentation technique used to address the problem of class imbalance in machine learning datasets; achieved 95.28%. Conversely, the hybrid ResNet50-PCA-Random Forest model achieved the highest classification accuracy of 98.43%.
On the other hand, for oil spill segmentation, three frameworks were developed: MobileNetV3–DeepLabV3, ResNet50–DeepLabV3 (where MobileNetV3 and ResNet50 serving as encoder backbones and an ASPP (Atrous Spatial Pyramid Pooling) segmentation module), and finally a Dual-Attention CNN. The experimental results showcase that MobileNetV3-DeepLabV3 model achieved the highest segmentation accuracy of 95.37% with an IoU of 90.90%, while DeepLabV3-ResNet50 achieved 95.21% accuracy with an IoU of 90.44%. The Dual-Attention CNN model achieved 83.09% accuracy. Overall, the proposed models demonstrate strong capability in oil spill classification and segmentation,providing reliable and efficient tools for environmental monitoring and decision support