← Projects

2025WorkML

Oil spill detection MVP

A computer vision MVP that detects oil spills in field images for an energy company, built on a fine-tuned YOLO detector.

Field images are labelled, preprocessed and used to fine-tune YOLO, evaluated against a zero-shot baseline and shown in a Gradio demo
From field images to a demo: labelled data fine-tunes YOLO, which is then compared against a zero-shot baseline.

The question

Can an energy company spot oil spills in field images automatically, and is a general-purpose model enough or does it need training?

What I did

  • Weighed zero-shot detection and segmentation against a trained detector, and chose YOLO.
  • Labelled the images and built their preprocessing.
  • Designed the MVP and built a Gradio demo to compare models side by side.

Results

The best fine-tuned model reached mAP50 0.95, with 0.93 precision and 0.90 recall.

Confidential client work: no screenshots, names or figures that identify the client.

  • Python
  • PyTorch
  • YOLO
  • Gradio
  • Docker
esc