Using AI to better detect and tackle oil spills in open seas and inland waterways

With the aim to explore how AI could improve oil spill detection and response management in the Netherlands, FruitPunch AI, Rijkswaterstaat and Valcon teamed up to design and develop a proof-of-concept.

Oil spills remain one of the most pressing environmental threats for the sea. The largest marine oil spill ever, the Deepwater Horizon slick in the Gulf of Mexico, spilled more than 134 million gallons of oil into the ocean.

Closer to home, oil spills are also a threat in the Netherlands too, particularly in the country’s ports and small waterways. These incidents pose risks to local ecosystems and demand swift action to minimise their impact. Traditional response methods – although effective – are often time-consuming, especially when a rapid cleanup can mean the difference between a minor incident and a major ecological disaster.

Recognising the urgent need for progress in the field, FruitPunch AI launched the ‘AI Against Oil Spills’ project, in collaboration with Rijkswaterstaat, the Dutch government body responsible for water management. The aim of the project was to try and understand how artificial intelligence could accelerate and improve oil spill detection and management in open seas, in ports and in inland waterways.

AI Against Oil Spills

To understand the opportunity for improvement, it is first key to understand the current way of working. Rijkswaterstaat currently deploys drone teams to capture aerial images of oil spills. Experts then manually review these images to estimate the volume of each spill, helping them determine the best cleanup strategy.

But this manual process is far from fast. While some inland spills can be assessed quickly using contextual information, spills at sea often require full segmentation of the affected area, an exercise that can take several hours. In a situation where time can be of critical importance, every minute lost is a missed opportunity to protect the environment and can spell the difference between a minor or major ecological disaster.

Valcon joined the ‘AI Against Oil Spills’ project, fielding a team of data experts to explore how they could potentially use AI for detecting oil spills and to improve the cleanup work at Rijkswaterstaat. The vision was to create an automated, AI-driven process that could replace manual inspection with real-time analysis, which could save hours of expert time and enable faster intervention.

A leading consulting firm with offices across Europe, Valcon developed a proof of concept to explore how to significantly reduce the time required to analyse drone imagery and estimate spill volumes. The proof of concept was developed through a three-stage approach:

Stage one: preparing the data
Like many real-world AI projects, the first hurdle was the data itself. The aerial images collected were not fully labelled and the scope of oil spills were often difficult to distinguish due to poor contrast or unclear boundaries. The Valcon team processed and edited the images to enhance the visibility of the oil, then manually re-labelled them to create a high-quality dataset suitable for machine learning.

Stage two: training the model
Next, Valcon trained several AI models to perform image segmentation, teaching the system to detect and outline oil spills from drone footage. The models learned to recognise patterns and textures indicative of oil, providing a visual map of the spill area in each image.

Stage three: estimating spill volume
The Valcon and Fruitpunch AI team developed a method to estimate spill volume based on the segmented images. By incorporating image metadata, such as drone altitude and camera settings, the AI could calculate a reliable volume estimate. Meaning that when a new drone image is uploaded, the AI tool can automatically outline the spill and calculate the volume within minutes.

The proof of concept

The proof of concept has demonstrated major potential for value added. The first major benefit is the speeding up of spill estimation. What currently takes hours could be done in minutes. The new AI-driven model provides Rijkswaterstaat with rapid volume estimates, enabling quicker and more targeted clean-up operations, which helps protect sensitive waterways before the damage spreads.

The second advantage of the proof of concept is the improvement of data collection. Valcon identified improvements to drone image collection that could boost the AI's accuracy, factors such as the shoreline comprising around 30% of the frame of each image, which could serve as reference points to locate the spill.

And while the model achieves around 70% accuracy in oil spill detection, it can help to guide the team, helping them make quicker decisions and hopefully reduce inspection time and enable faster environmental response.

AI to the use

“We are delighted to have supported FruitPunch AI and Rijkswaterstaat with the proof of concept. It showcases the powerful role AI could play in solving real-world environmental challenges, and how technology and human expertise can come together to co-create unique innovations.”