This repository contains implementations and evaluations of three self-supervised denoising models for enhancing low-cost satellite imagery.
- Filippo Balzarini - filippo.balzarini@mail.polimi.it - Politecnico di Milano
- Michele Cavicchioli - michele.cavicchioli@mail.polimi.it - Politecnico di Milano
This project investigates whether imagery acquired from hypothetical low-cost satellites (characterized by spatially correlated noise) can be made suitable for land monitoring tasks through modern denoising techniques. The evaluation focuses on three state-of-the-art architectures based on Blind-Spot Networks (BSN).
Repository: AP-BSN Implementation
A self-supervised denoising approach that uses pixel-shuffle downsampling to generate multiple patches with reduced noise correlation. The asymmetric architecture applies blind-spot constraints to these downsampled images.
Repository: NL-N2V Implementation
Extends the original Noise2Void by introducing a non-local masking strategy, replacing masked regions with similar distant patches. Particularly effective for spatially correlated noise.
Repository: SSID Implementation
Combines a blind-neighborhood network for flat regions with a locally aware network for textured areas, weighted by local flatness. This adaptive approach balances noise removal across different image characteristics.
The evaluation used Sentinel-2 satellite imagery from multiple urban locations:
- Rome, Paris, Madrid, Tokyo, London, Milan
- Acquisition dates: 2023-2024
- Bands: B04 (Red), B03 (Green), B02 (Blue)
- Total images: 3,979 patches (512×512 pixels)
- Split: 80% training, 10% validation, 10% testing
A two-stage synthetic noise injection process simulates low-cost satellite conditions:
- Poisson (Photon) Noise: Models photon-limited sensor noise
- Gaussian (Electronic) Noise: Band-dependent electronic noise with spatial correlation
Parameters:
- Photon scale: 1000
- Noise boost: 10.0
- Gaussian kernel: 5×5, σ = 0.55
- SNR values: {230, 249, 214} for R, G, B bands (official Copernicus calibration report)
Noise injection procedure
- SSIM (Structural Similarity Index): Measures structural preservation and perceptual quality
- PSNR (Peak Signal-to-Noise Ratio): Measures pixel-level reconstruction accuracy
| Model | SSIM ↑ | PSNR ↑ | Patch Size |
|---|---|---|---|
| APBSN | 0.795 | 30.71 | 480x480 |
| NLN2V | 0.787 | 28.56 | 240x240 |
| NLN2V | 0.789 | 28.45 | 512x512 |
| SSID | 0.799 | 29.92 | 512x512 |
✅ Strengths:
- All models effectively reduce noise levels
- Suitable for large-scale structure analysis (agricultural fields, major buildings)
- Significant loss of fine details across all models
- Reduced performance in low-contrast scenarios
- Challenges in dense urban environments
- Blind-spot mechanisms inherently suppress small-scale textures