Oil Spill Detection from Satellite SAR Images using CNN

When oil hits the sea surface it damps the tiny wind-driven capillary waves that radar reflects off, so spills appear as dark patches in synthetic-aperture-radar images — visible day and night, through cloud cover. The catch is that low-wind zones, biogenic films and other look-alikes produce the same dark signature, and telling them apart is the whole problem. This project takes the modern approach: a U-Net style encoder-decoder CNN trained for five-class semantic segmentation on the MKLab oil-spill benchmark, learning shape, texture and context instead of relying on hand-crafted dark-spot features. The training notebook covers SNAP-style preprocessing, augmentation, weighted loss functions for the class imbalance, and honest per-class IoU reporting. A web demo lets you pick real SAR scenes — including the TerraSAR-X Montara and Gulf of Mexico spill imagery — and watch the detector classify each scene with per-class probabilities and a segmentation overlay.

1. Sentinel-1 VV-polarization SAR scenes are preprocessed (region cropping, radiometric calibration, speckle filtering, dB-to-luminance conversion) following the benchmark's published pipeline.
2. The 1,112 annotated scenes (1,002 train / 110 test) provide five-class masks: sea surface, oil spill, look-alike, ship and land.
3. A U-Net encoder-decoder is trained with weighted cross-entropy (plus focal-loss experiments) to handle the severe class imbalance — spill pixels are rare.
4. Augmentation (rotation, flipping) expands the effective training set, since SAR scenes are robust to affine transforms.
5. The test split is evaluated once: per-class IoU, with the oil-spill class as the headline metric, plus qualitative overlay inspection.
6. In the web demo, a selected scene passes through the saved model; the app shows the five class probabilities, the verdict, and the toggleable segmentation overlay.

Tech stack:

• Python 3, PyTorch (or TensorFlow/Keras — buyer's choice)
• NumPy, Matplotlib, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (detection-studio web demo)
• SNAP-style SAR preprocessing chain
• MKLab oil-spill SAR benchmark
• Git

Dataset & model details

• Dataset: MKLab oil-spill SAR benchmark (Krestenitis et al., "Oil Spill Identification from Satellite Images Using Deep Neural Networks", Remote Sensing 11(15), 2019) — 1,112 Sentinel-1 C-band SAR scenes (VV polarization) at 1250×650 px, fixed 1002/110 train/test split. Events span 28 Sep 2015 – 31 Oct 2017, confirmed by EMSA CleanSeaNet. RGB visualisation masks plus single-channel integer masks (0–4) are provided.
• Task: 5-class semantic segmentation; input = 1250×650×1 SAR scene, output = per-pixel class map over {sea surface, oil spill, look-alike, ship, land}.
• Model: U-Net encoder-decoder (contracting path + expansive path with skip connections); weighted cross-entropy loss with focal-loss variant; rotation/flip augmentation.
• Metrics: Per-class IoU; reported literature baselines are ≈0.54 IoU on the oil-spill class for U-Net/DeepLabv3+ (Krestenitis et al., 2019). The built-to-order training run targets comparable performance — stated as a design target, with the report documenting the actual achieved figures after training. No accuracy is claimed as measured until the training run is executed for the order.

• [U-Net segmentation model] Encoder-decoder CNN with skip connections for five-class pixel-wise segmentation — sea surface, oil spill, look-alike, ship and land.
• [Real SAR demo scenes] The web demo runs over genuine TerraSAR-X radar imagery (Montara platform spill, Gulf of Mexico oil-polluted area, clean open sea) with per-class probability bars.
• [Segmentation overlay] Toggleable mask overlay on each scene showing where the model places the spill boundary.
• [Full training notebook] Data loading, augmentation, model definition, weighted cross-entropy and focal loss, training loop and evaluation in one reproducible notebook.
• [Per-class IoU reporting] Intersection-over-union for every class, with explicit analysis of the oil-spill vs look-alike confusion that defines this task.
• [Look-alike analysis] Dedicated evaluation of the hardest distinction in the field — dark patches that are not oil — documented for the viva.
• [Exported trained model] Saved weights plus the preprocessing pipeline, so the demo runs the real network without retraining.

• Complete training & evaluation Jupyter notebook
• Trained segmentation model weights with preprocessing code
• Detection-studio web demo with real SAR scenes, probability bars and overlay toggle
• Per-class IoU report and look-alike confusion analysis
• Project report PDF (SAR principles, dataset, architecture rationale, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (SAR backscatter, Doppler-free imaging, U-Net, IoU, class imbalance, look-alikes)