NeuralNomad0101/sar2eo-pix2pix

SAR-to-EO Image Translation — Pix2Pix Generator

U-Net generator trained on the SEN1-2 dataset for translating Sentinel-1 SAR (VV) patches into Sentinel-2 EO (RGB) imagery. Part of a GalaxEye Satellite AI Research Intern assessment.

Model Description

Property	Value
Architecture	U-Net (8 encoder levels, 7 decoder + final)
Parameters	~55M
Input	1-channel SAR VV, 256×256, normalized to [-1, 1]
Output	3-channel RGB, 256×256, Tanh activation in [-1, 1]
Training data	SEN1-2 (16,000 paired patches, 4 terrain classes)
Training duration	200 epochs
Loss	L1 (λ=100) + adversarial (BCE)

Usage

import torch
from models import UNetGenerator

G = UNetGenerator(in_channels=1, out_channels=3)
state = torch.load("gen_best.pth", map_location="cuda", weights_only=True)
G.load_state_dict(state)
G.eval()

# SAR tensor: [1, 1, 256, 256] normalized to [-1, 1]
with torch.no_grad():
    eo = G(sar_tensor)  # [1, 3, 256, 256] in [-1, 1]

Performance

Metric	Validation	Test
LPIPS ↓	0.3824	0.3848
FID ↓	94.83	96.33
SSIM ↑	0.2768	0.2740
PSNR ↑ (dB)	17.76	17.59

Ablation

Compared against an L1-only baseline (same U-Net, no discriminator). The adversarial loss improves LPIPS by −34% and FID by −42%, at the cost of pixel-level accuracy (SSIM −35%, PSNR −13%).

Files

• gen_best.pth — Generator weights (best by validation L1, epoch 180)

References

• Isola et al., "Image-to-Image Translation with Conditional Adversarial Networks" (Pix2Pix), CVPR 2017
• Schmitt et al., "SEN1-2: A Dataset of Sentinel-1 and Sentinel-2 Image Pairs", 2018