Utilizing deep learning algorithms for automated oil spill detection in medium resolution optical imagery.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Environmental Monitoring

Background:

• Oil spills pose significant environmental threats, necessitating effective detection methods.
• Optical satellite imagery offers a valuable resource for monitoring marine environments.
• Deep learning algorithms show promise for automated analysis of satellite data.

Purpose of the Study:

• To evaluate and enhance deep learning models for oil spill detection using Sentinel-2, Landsat-8, and Landsat-9 imagery.
• To investigate the impact of attention mechanisms on the performance of convolutional neural network architectures.
• To identify the optimal deep learning model for accurate oil spill detection.

Main Methods:

• Developed a semi-automatic labeling approach for creating oil slick datasets from global spill cases.
• Enhanced UNet, BiSeNetV2, and DeepLabV3+ architectures with attention modules (SE, CBAM, SimAM).
• Evaluated model performance using medium-resolution optical satellite imagery.

Main Results:

• The UNet architecture integrated with the Convolutional Block Attention Module (CBAM) demonstrated superior performance.
• Incorporating sun glint as a feature significantly improved the UNet-CBAM model's accuracy.
• The optimal UNet-CBAM model achieved a micro-average F1 score of 88.8% for oil spill detection.

Conclusions:

• Deep learning, particularly attention-enhanced UNet, shows significant potential for oil spill detection in optical remote sensing.
• The findings underscore the increasing relevance of these technologies with the proliferation of high-resolution satellite data.
• Further research can leverage these advanced models for more effective environmental monitoring and response.