osama-bin/sar-oil-spill-classifier

SAR Oil Spill Thickness Classifier

A CNN + Swin Transformer hybrid deep learning model that classifies oil spill thickness from Sentinel-1 SAR (Synthetic Aperture Radar) images.

What it does

Upload a Sentinel-1 SAR image and the model will:

1. Detect oil presence in the SAR image
2. Classify the oil thickness into one of three categories:
   • Thin_Sheen -- thin oil-sheen-like SAR signature
   • Moderate -- intermediate thickness / emulsion
   • Thick_Emulsified -- thick or emulsified oil
3. Return confidence scores for each class

Model Architecture

Component	Details
CNN Branch	ResNet-18 backbone (512-dim features)
Swin Branch	Swin-Tiny patch4 window7 224x224 (768-dim features)
Fusion	Concatenation (1280-dim) -> Linear -> BatchNorm -> ReLU (512-dim)
Output	Softmax -> 3-class probability distribution

Usage

Upload any Sentinel-1 SAR image and click Analyse SAR Image.