Research shows how AI models can help by making environmental monitoring faster and cheaper.

By Sigurd Almli Hanssen, Doctoral Research Fellow at SFI Visual Intelligence

Marine oil spills are a serious threat to marine life and coastal communities. They spread quickly and keep drifting with wind and currents, covering ever larger areas. Limiting the damage means detecting spills quickly, which requires monitoring the ocean surface at high frequency over massive areas. This is exactly where AI models help, making environmental monitoring faster and cheaper.

Satellite images are the most cost effective way to monitor areas this large. Optical images are more accurate but need daylight and clear skies. SAR images are the other option, radar images that see through cloud cover and darkness, especially useful in places like northern Norway. The catch is that you're not directly observing an oil spill, but its signature, since spills smooth out the small waves wind creates on the surface.

Unfortunately, ship wakes, wind shadows behind platforms or islands, and algae blooms all create a similar signature, making lookalikes hard to tell apart from real spills. The problem is that labeled examples of oil spills are relatively rare in public datasets, and lookalike examples are especially rare, since unlike oil spill labels, they aren't a byproduct of any commercial detection service.

To get around this, we reframed the task. Instead of teaching the model what an oil spill looks like, we teach it what clean sea looks like, using a diffusion model for anomaly detection, a method already used elsewhere, like spotting brain tumors in medical scans, where normal data is plentiful and anomalies rare. These models are trained to remove noise added to training images, so the idea is that it gets good at restoring normal sea conditions but struggles with anything it hasn't seen before, mostly oil spills, since that's what we deliberately left out of training. From there, we can flag anomalies, meaning oil spills, based on how much the model struggles to reconstruct each image.

We tested this on both a public dataset and a real operational dataset from KSAT, comparing it against a simpler model called an autoencoder. The diffusion model did a solid job separating clean sea from anything unusual, oil spill or lookalike alike, and two tricks made this even better. Multiplying the anomaly map by the negative of the image highlighted error in dark areas like oil spills, and running the model several times, up to five, then taking the maximum error at each pixel, made results more accurate and stable. Still, the model struggled to tell oil spills and lookalikes apart from each other, which suggests the two just aren't distinct enough for it to separate well.

This is really just an initial step toward turning oil spill detection into an unsupervised task, but it's a useful one. It shows this kind of approach can work as a first filter, flagging which patches of ocean are worth a closer look without needing a single labeled oil spill to get started. Actually telling spills and lookalikes apart will probably need a smarter method that more actively tries to distinguish the two classes.

Pending accept in renowned journal. More information coming soon.