Binary Change Detection (EO-SAR Fusion)

This model is a Siamese UNet designed for binary change detection using fused Electro-Optical (EO) and Synthetic Aperture Radar (SAR) satellite imagery.

Model Description

The architecture uses dual weight-shared ResNet-34 encoders to extract multi-modal features from pre-event RGB (EO) and post-event grayscale (SAR) images. Feature differences are fused via skip connections into a UNet decoder.

• Encoder: ResNet-34 (Pre-trained on ImageNet)
• Weights Format: Safetensors
• Loss Function: Combined Focal Loss + Dice Loss
• Input Resolution: 256x256

Training Results (Final Run)

• Best Validation IoU: 24.74%
• Precision: 26.45%
• Recall: 79.29%
• F1 Score: 39.67%

How to Use

To use this model, ensure you have the src/ folder from the GitHub repository in your local directory.

import torch
from safetensors.torch import load_file
from src.model import SiameseUNet

# 1. Initialize architecture
model = SiameseUNet(pretrained=False)

# 2. Load professional weights
weights = load_file("model.safetensors")
model.load_state_dict(weights)

# 3. Ready for inference
model.eval()