Situation
Oil spills pose significant threats to our ecosystems and unfortunately, they are a frequent and damaging occurrence in the ports and small waterways of the Netherlands. These spills demand prompt and efficient action to mitigate environmental harm.
‘AI for good’, an initiative organised by Fruitpunch AI, has been exploring the potential for using AI for environmental good. This includes a collaborated project with Rijkswaterstaat, a Dutch government organisation responsible for water management, to understand how AI can help identify, manage and resolve oil spills in Dutch ports.
Currently, Rijkswaterstaat deploys drone teams to capture aerial images of spills, which experts manually inspect to estimate the spill’s volume—a critical step in determining the appropriate clean-up strategy. While this method is reliable, it is time consuming. Spills on inland waterways can often be quickly managed using auxiliary information, but spills at sea require full area segmentation and can take several hours to calculate the spill volume. These delays highlight the need for faster, more efficient solutions.
In response to this challenge, Valcon fielded a team of data experts to participate in the ‘AI against oil spills’ project aimed at accelerating the oil spill response process.
Approach
Valcon proposed a solution to significantly reduce the time required for spill volume calculations. Our goal was to develop an AI-driven solution that could automate the manual inspection of drone images.
Valcon followed a three-step approach. The first was data preparation and preparing the aerial image data for training – the data set presented challenges including incomplete labelling, so the images were processed and edited to enhance oil visibility and manually re-labelled the oil spills.
The second was to train several AI models for image segmentation and build a model to detect and outline oil spills. The third was spill volume estimation and developing a method to estimate spill volume based on image metadata and the type of oil being spilled. Altogether, when a new drone image is fed into the built model, the AI draws boundaries around the detected oil spill, which combines with the additional information of oil spill type to estimate the spill volume.
Results
Faster spill volume estimation: Valcon devised a method to estimate the volume of the oil spills from drone images within a few minutes. This serve as a timely and helpful resource for Rijkswaterstaat in implementing a targeted cleanup response.
Improving data quality: Valcon identified improvements to drone image collection that could boost the AI’s accuracy. For example, we ensured that the shoreline comprises approximately 30% of the frame of each image to serve as reference points. We provided these insights to the drone team, who could immediately apply them in the field.
Detection accuracy: although the model’s detection accuracy is currently around 70% and there is a room for improvement, it can still provide useful input for experts, helping them make quicker decisions and potentially shortening the overall time spent on oil spill inspections.
