Oil spills, lookalikes and a whole lot of noise: A diffusion based approach for Oil Spill Detection in SAR Images

Abstract

Accurate and fast oil spill detection is crucial to enable effective usage of mitigation resources to minimize environmental damage. Synthetic Aperture Radar (SAR) imagery provides a reliable means of large-scale ocean monitoring, but analyzing these images remains a time-consuming and labor-intensive task. The presence of lookalikes in SAR images, natural phenomena imitating the appearance of oil spills, provide a potential source of false alarms for both human analysts and automated methods.

This thesis explores the use of reconstruction-based models for detecting oil spills. We reframe the oil spill detection problem as an Out-of-Distribution detection task by training the reconstruction model exclusively on images without oil spills, where anomaly maps can be created from the reconstruction errors. A diffusion model is implemented and compared to a standard autoencoder and a classical baseline texture descriptor using Local Binary Patterns (LBP). All models' anomaly (or LBP) maps are converted to histograms and classified using a Support Vector Machine. To the best of our knowledge, this is one of the first studies applying diffusion models to SAR images specifically for oil spill detection.

While recent work has shown diffusion models to be especially useful for anomaly detection tasks, this study highlights the challenges in adapting diffusion models to SAR imagery: The combination of the fine-grained features of SAR and the need to balance noise levels in the diffusion process made detection of small or diffuse lookalikes particularly difficult. The autoencoder outperformed the diffusion model with an similar recall, both at 70% but with a significantly higher precision at 59%. The LBP on the other hand achieved a very poor precision, but strong recall.

This study has shown some of the potential and highlighted the main limitations related to diffusion-based approaches in SAR image analysis. Multiple directions for future work have been identified, including hyperparameter tuning of the noise level and conditioning variable, and dataset expansion and refinement.