Skip to content

s1_sr_concept.md

Latest commit

 

History

History
47 lines (32 loc) · 3.12 KB

s1_sr_concept.md

File metadata and controls

47 lines (32 loc) · 3.12 KB

Concept: Sentinel-1 Super-Resolution & Despeckling Integration

Executive Summary

Unlike Sentinel-2, where super-resolution is primarily an optical upscaling task, Sentinel-1 (SAR) super-resolution is intrinsically linked with Speckle Reduction. This concept proposes integrating a deep-learning-based Super-Resolution (SR) stage into the S1 pipeline to enhance Ground Range Detected (GRD) products from 10m to 2.5m or 5m.

1. Technical Approach: Joint SR and Despeckling

Standard interpolation (Bilinear/Lanczos) in SAR data amplifies speckle noise, leading to "grainy" high-res images. The optimal integration uses models that perform Joint Super-Resolution and Despeckling.

Potential Libraries

  • Solafune-Tools: Provides a 5x SR model for Sentinel imagery.
  • THREASURE-Net: Uses EDSR (Enhanced Deep Residual Networks) to reach 2.5m resolution, specifically optimized for Sentinel-1/2 time series.
  • Real-ESRGAN / Custom CNNs: Recent research (e.g., Ayala et al. 2024) suggests using S1 Stripmap (SM) mode data as a training target to upscale standard Interferometric Wide (IW) mode data.

2. Integration Architecture

A. Strategic Placement in functions_s1.py

The SR stage must occur after calibration but before denoising/warping, or as a replacement for the current denoise.py stage.

  1. Calibration: Perform radiometric calibration as usual (s1_calibrator.py).
  2. SR Stage (New): Apply the SR model to the calibrated (but still in slant/ground range) data.
  3. Warping: Use gpu_warp.py to project the high-res 2.5m data to EPSG:3857.

B. Impact on Products

Product Impact
VV / VH (Visual) Significant improvement in feature edge definition (coastlines, large vessels, urban structures).
RATIO VV/VH Cleaner ratio maps with less "salt and pepper" noise, improving target detection.
Fusion (TARGET-PROBE) 2.5m SAR data aligned with 2.5m S2 data (from SEN2SR) creates an ultra-high-resolution multi-sensor composite.

3. Implementation Challenges

  • Compute Intensity: SAR SR is computationally heavier than optical SR due to the need to handle complex signal characteristics or high-dynamic-range float data.
  • Artifacts: Over-smoothing can occur with GAN-based models, potentially hiding small tactical targets (e.g., small vehicles or masts) while making the overall image look "prettier."
  • Data Size: A 2.5m S1 tile is 16x larger than a 10m tile.

4. Proposed Strategy: "The High-Res Fusion"

The most powerful use case for S1 SR in this pipeline is to match the resolution of the proposed S2 SEN2SR integration.

  1. Enable S1_SUPER_RESOLVE=true in .env.
  2. Use a model like EDSR or Swin-Transformer to reach 2.5m.
  3. Directly feed the 2.5m SAR data into the correlate.py engine to create 2.5m Fused Products.

5. Conclusion

While more scientifically complex than S2 upscaling, S1 Super-Resolution is the "missing link" for a true high-resolution monitoring pipeline. Integrating a joint despeckling/SR model would elevate the pipeline's output to a level comparable with commercial high-res SAR providers.