Research question
Can GAN-based super-resolution outperform CNN baselines on Sentinel-2 multispectral imagery at 4× upscale — and if so, at what perceptual cost?
Dataset
Sentinel-2 L2A imagery over mixed urban/agricultural areas. Pairs of:
- Low-resolution input: 10m/px bands (B2, B3, B4, B8) downsampled to 40m/px
- High-resolution target: original 10m/px
~12,000 patches of 64×64 px (LR), 256×256 px (HR).
Models compared
| Model | Architecture | Loss | |---|---|---| | SRCNN | 3-layer CNN | MSE | | RCAN | Residual channel attention | MSE | | SRGAN | ResNet generator + PatchGAN discriminator | MSE + adversarial + perceptual | | ESRGAN | Enhanced SRGAN with RRDB blocks | L1 + adversarial + perceptual |
Key finding
GANs (SRGAN, ESRGAN) consistently produced sharper-looking output that fooled visual inspection. PSNR and SSIM told a different story: the CNN baselines scored higher on both metrics. The adversarial loss was introducing high-frequency texture that looked plausible but wasn't faithful to the ground truth.
For applications where geometric accuracy matters (crop boundary detection, building footprint extraction), this is a real problem. For visual presentation, GANs won.
Infrastructure
Full PyTorch training pipeline with:
- Reproducible data pipeline (deterministic augmentation, fixed seeds)
- Mixed-precision training (AMP) for 2× throughput on the lab GPU
- MLflow experiment tracking
- Evaluation harness computing PSNR, SSIM, LPIPS, and FID per model