Image Super-Resolution using Generative Models on the SEN2VENµS Dataset
Qualitative results of super-resolution on multispectral satellite imagery.

Image Super-Resolution using Generative Models on the SEN2VENµS Dataset

Author

  • Debwashis Borman

Dataset

I used the SEN2VENµS dataset, which is designed specifically for super-resolution in remote sensing.
It provides paired Sentinel-2 and VENµS observations across 29 global sites.

For my experiments, I picked 6 representative sites:

  • ALSACE (France, Alsace)
  • ANJI (China, Zhejiang Sheng)
  • BENGA (India, West Bengal)
  • LERIDA-1 (Spain, Catalonia)
  • NARYN (Kyrgyzstan, Naryn)
  • SO2 (France, Midi-Pyrénées)

To make the evaluation fair, I split the data by site (so there’s no spatial overlap):

  • Train: 12,355 pairs (~70%)
  • Validation: 2,649 pairs (~15%)
  • Test: 2,650 pairs (~15%)

Tasks I worked on:

  • 2× Super-Resolution
    • Input: Sentinel-2 RGBI bands (10m)
    • Target: VENµS (5m)
  • 4× Super-Resolution
    • Input: Sentinel-2 Red Edge bands (20m)
    • Target: VENµS (5m)

Data Visualization Data Visualization after extraction and Processing

Models

I compared both classical deep SR models and GAN-based approaches:

  • SRCNN
    The first CNN for super-resolution. Very simple 3-layer design – works better than bicubic, but tends to blur fine details.

  • RCAN
    A much deeper CNN with attention modules. This one is great at preserving structure and sharpness, and usually gives the best scores.

  • ESRGAN
    A GAN-based model that focuses on realism and textures. It uses Residual-in-Residual Dense Blocks (RRDB) and a perceptual loss based on VGG19 features.
    It often produces visually sharper results, though sometimes at the cost of pixel-accuracy.

ESRGAN Architecture ESRGAN Architecture

Results

I evaluated the models in two ways:

  • Quantitatively using PSNR, SSIM, and LPIPS
  • Qualitatively by visually comparing cropped regions

Quantitative Results

2× SR (Sentinel-2 RGBI → VENµS 5m)

Model PSNR SSIM LPIPS (RGB) LPIPS (NIRG)
SRCNN 38.39 0.960 0.095 0.133
RCAN 42.32 0.981 0.056 0.079
ESRGAN 37.45 0.950 0.085 0.152

4× SR (Sentinel-2 Red Edge → VENµS 5m)

Model PSNR SSIM LPIPS (567) LPIPS (678)
SRCNN 34.71 0.903 0.294 0.248
RCAN 37.36 0.936 0.178 0.155
ESRGAN 37.51 0.928 0.177 0.154
ESRGAN (8ch) 36.38 0.925 0.279 0.345

Qualitative Results

2× RGB bands:

2x RGB Results HR with ROI (left) and crops from HR, LR, SRCNN, RCAN, ESRGAN (right).

2× NIRG bands:

2x NIRG Results Comparison on NIRG bands.

4× Red Edge (5-6-7 bands):

4x Red Edge Results Comparison on 5-6-7 bands.

Observations

  • RCAN is the most reliable when it comes to numerical accuracy (best PSNR/SSIM and lowest LPIPS).
  • ESRGAN produces sharper and more realistic textures, making the images “look” better, even if the scores are a bit lower.
  • SRCNN is a solid baseline, but clearly outdated compared to the other two.

Final Note

This project was a great learning experience – from handling multispectral remote sensing data to adapting GANs for SR tasks.
As the sole author, I built everything from dataset preparation to model training and evaluation, and I’m excited to share the results here.