This repository presents project Nafta, which focuses on oil spill detection and segmentation. In addition, YOLOv3 model is implemented to detect oil rigs and vessels. All models trained on the Nafta dataset can be downloaded [here][Models]. Please read the following article for more detailed information about the project and procedures.
Please cite this article if you use the dataset, model or method(s), or find the article useful in your research:
Monitoring offshore oil pollution using multi-class convolutional neural networks
Citation: Ghorbani Z., Behzadan A.H. (2021), “Monitoring Offshore Oil Pollution Using Multi-Class CNNs”, Environmental Pollution, 289: 117884.
The Nafta dataset contains 1,292 images. In total, 70% of images are taken by drones (low altitude), 16% by satellites (high altitude), and the remaining 14% from first-person views. There are 1,431 instances of the “oil spill” class, 959 instances of the “vessel” class, and 316 instances of the “oil rig” class (See the Venn Diagram Below). The dataset is available here.
VGG16 architecture is modified and fine tuned for oil spill classification. You can download the trained model from the "VGG16" subfolder under the "Models" folder in the dataset.
Pyramid Scene Parsing Network (PSPNet) architecture is modified and fine tuned for oil spill semantic segmentation. You can download the trained model from the "PSPNet" subfolder under the "Models" folder in the dataset. Grondtruth masks for the oil spill class can be found in the "Masks" folder in the dataset.
Mask R-CNN architecture is modified and fine tuned for oil spill instance segmentation. You can download the trained model from the "Mask R-CNN" subfolder under the "Models" folder in the dataset. Grondtruth masks for the oil spill class can be found in the "Masks" folder in the dataset.
YOLOv3 model is used for vessel and oil rig detection. You can download the trained model from the "YOLOv3" subfolder under the "Models" folder in the dataset. The annotations can be found in the "Labels" folder in the dataset.
