GalaxEye Binary Change Detection (EO-SAR)

Project Title & Description

Siamese UNet for Imbalanced Binary Change Detection using EO-SAR Images

This repository contains the solution for the GalaxEye Satellite AI Research Intern assessment. The task involves performing pixel-level binary change detection (0 = No-Change, 1 = Change) given co-registered pre-event (EO) and post-event (SAR) image pairs.

The solution implements a Siamese UNet architecture with weight-shared ResNet-34 encoders to extract multi-modal features, absolute difference-based skip fusion, and a combined BCE+Dice loss function to tackle severe class imbalance (the dataset exhibits a 58:1 ratio of no-change to change pixels).

Requirements

• Python 3.11+
• See requirements.txt for all pinned dependencies.

Environment Setup

Run the following commands to create a virtual environment and install dependencies:

# 1. Create a virtual environment
python3 -m venv .venv

# 2. Activate it
# On Mac/Linux:
source .venv/bin/activate
# On Windows:
# .venv\Scripts\activate

# 3. Install requirements
pip install -r requirements.txt

Dataset Structure

Place the provided dataset inside a folder named Dataset/ at the root of the repository. The expected directory layout is:

Galaxyeyeai/
├── Dataset/
│   ├── train/
│   │   ├── pre-event/
│   │   ├── post-event/
│   │   └── target/
│   ├── val/
│   │   ├── pre-event/
│   │   ├── post-event/
│   │   └── target/
│   └── test/
│       ├── pre-event/
│       ├── post-event/
│       └── target/
├── src/
├── config.yaml
├── train.py
├── eval.py
└── ...

Note: The script automatically handles the mandatory label remapping (0/1 -> 0, 2/3 -> 1).

Training

To train the model from scratch, execute:

python train.py --config config.yaml

Checkpoints will be saved automatically to checkpoints/best.pth and checkpoints/last.pth. TensorBoard logs are stored in runs/.

Evaluation

To evaluate the model on the test data (or any split) and generate metrics + visualisations, run:

python eval.py --data_path Dataset/test --weights checkpoints/best.pth

Results will be saved inside the results/ folder, including JSON metrics, confusion matrices, and qualitative image comparisons.

Model Weights

(Insert public link to your model weights here. e.g., Google Drive or Hugging Face Hub link)

• Link: Hugging Face Repository

Results

These metrics are placeholders, assuming the model has finished training on local hardware.

Split	IoU	Precision	Recall	F1 Score
Validation	24.74%	26.45%	79.29%	39.67%
Test	3.50%	4.19%	17.37%	6.75%

Citation / References

• UNet: Ronneberger et al., "U-Net: Convolutional Networks for Biomedical Image Segmentation", MICCAI 2015.
• Siamese Networks for CD: Daudt et al., "Fully Convolutional Siamese Networks for Change Detection", ICIP 2018.
• Albumentations: Buslaev et al., "Albumentations: fast and flexible image augmentations", Information 2020.
• Timm: Ross Wightman, "PyTorch Image Models", GitHub.