OceanWatch Models
Trained deep-learning models used by OceanWatch, a maritime oil-spill detection and vessel-attribution system for Indian waters.
GitHub: https://github.com/ThatAnacondaGuy/OceanSpill
Models
| Model | Task | Input | Format |
|---|---|---|---|
| SAR Oil | Oil segmentation | Sentinel-1 VV + VH | ONNX |
| SAR Single-Pol | Oil segmentation | Sentinel-1 VV | ONNX |
| SAR Smoke | SAR segmentation experiment | Sentinel-1 VV + VH | ONNX |
| Optical | Oil segmentation | Sentinel-2 optical | ONNX |
| Ship Detection | Vessel segmentation | SAR | ONNX |
SAR Oil
Architecture: U-Net
Input: 2-channel Sentinel-1 SAR
Channels: VV + VH
Loss: Dice + Focal Loss
SAR Single-Pol
Architecture: U-Net
Input: 1-channel Sentinel-1 SAR
Channel: VV
This model was trained to evaluate whether a single-polarisation input was sufficient for oil-spill segmentation.
SAR Smoke
Architecture: U-Net
Input: 2-channel Sentinel-1 SAR
Channels: VV + VH
Validation metrics:
- IoU: 0.5401
- Dice: 0.7014
- Precision: 0.8095
- Recall: 0.6188
- False alarm rate: 12.28%
Optical
Architecture: U-Net
Input: Sentinel-2 optical imagery
Channels: Blue, Green, Red, NIR
Ship Detection
Architecture: U-Net
Task: Vessel detection/segmentation in SAR imagery
Model Files
Each model directory contains:
unet_best.onnx— ONNX model graphunet_best.onnx.data— external ONNX tensor dataunet_best.json— training and evaluation metadata
The .onnx and .onnx.data files must be kept together.
Intended Use
These models are research/prototype models developed for the OceanWatch project.
They are intended for:
- oil-spill detection research
- maritime remote sensing
- vessel detection research
- experimentation with satellite imagery
- hackathon demonstration
They should not be treated as definitive evidence of an oil spill or vessel involvement without independent verification.
Limitations
Model performance depends on the imagery, preprocessing, sensor characteristics, environmental conditions and threshold used.
Detection of a dark region in SAR imagery does not by itself establish that the region is oil.
Vessel detections and attribution outputs require additional contextual and observational evidence.
Project
OceanWatch was developed for maritime oil-spill detection and vessel attribution.
Source code: https://github.com/ThatAnacondaGuy/OceanSpill