juggtimber/galaxeye-eo-sar-change-detection

GalaxEye EO--SAR Change Detection

Lightweight multimodal EO--SAR change detection framework based on EfficientNet-B0 UNet architecture.

Overview

This model performs binary pixel-level change detection using paired Electro-Optical (EO) and Synthetic Aperture Radar (SAR) imagery.

The framework combines:

• EO RGB channels
• CLAHE-enhanced SAR
• Log-transformed SAR

into a 5-channel multimodal segmentation input.

Architecture

• EfficientNet-B0 encoder
• UNet-style decoder
• BCE + Dice + Focal hybrid loss

Results

Split	IoU	F1
Validation	0.2712	0.4267
Test	0.0887	0.1629

Limitations

Performance degrades under cross-scene EO--SAR domain shift and heterogeneous SAR backscatter conditions.

Repository

GitHub repository: https://github.com/timbersaw-jugg/galaxeye-eo-sar-change-detection/