Case study
The hard class is oil, not background. The evaluation therefore has to reward finding oil instead of hiding behind headline pixel accuracy.
Oil slicks appear as dark patches in Sentinel-1 SAR, but low-wind zones, biogenic films and rain cells can look almost identical. The technical problem is distinguishing oil from those look-alikes while carrying the result from a trained model into georeferenced, usable output.
Role & scope
A five-class semantic-segmentation system spanning training and model comparison, oil-specific evaluation, ONNX export, Sentinel-1 ingestion/preprocessing, tiled inference, vectorized geospatial output, FastAPI serving and a React/MapLibre interface.
Approach
How the system earns the result.
01 — Frame
Model the confusing classes explicitly.
The task uses five classes — sea surface, oil spill, look-alike, ship and land — so look-alike confusion is represented rather than folded into background.
02 — Compare
Select on oil-specific evidence.
U-Net, DeepLabV3+ and SegFormer are evaluated on the committed test split. SegFormer mit-b2 is selected using oil IoU, oil recall and broader per-class metrics rather than background-dominated accuracy.
03 — Operationalize
Move from chips to full georeferenced scenes.
The selected model is exported to ONNX, reused for tiled scene inference and converted into GeoTIFF plus oil polygons before API delivery.
04 — Challenge
Run the pipeline on a real unseen event.
The public Wakashio case study applies the end-to-end pipeline to a Sentinel-1B scene over Mauritius and documents both the detected location and the limitations of measured area.
Evidence
What can actually be checked.
Oil IoU
0.566
Selected SegFormer mit-b2 result on the committed 110-image test split.
Oil recall
0.764
Oil-class recall from the same committed evaluation output.
Mean IoU
0.696
Five-class mean IoU; reported alongside, not instead of, oil-specific quality.
Real-event test
Wakashio
A previously unseen August 2020 Sentinel-1B scene over the MV Wakashio spill is included as a public case study.
Limits & boundaries
What the case study does not pretend.
The official dataset is small and imbalanced: 1002 training images, 110 test images, with oil occupying roughly 1% of pixels.
Training uses a single VV SAR channel; dual-polarization information is not exploited.
The raw-scene case study crosses a preprocessing/domain gap, so detected area is presented as an approximate lower bound rather than false measurement certainty.
Publication boundary
The modernized repository explicitly avoids using the original project's background-dominated accuracy as the headline result. The case study preserves that same metric discipline.
From proof to useful work
Where this project maps to real service work.
These links come from the governed project/service evidence map. They are not generic cross-sells and do not widen the claims made above.
Next project · Geospatial decision system
Solar Site Selection