Home
ES · Ver en español

Education project

Sentinel-1 Radar Imaging

From radar noise to a dataset, and two models that did not beat the input

Ten radar images of a coastal area, filtered, clustered with K-Means and turned into a dataset, plus two restoration models that could not improve on the input.

Three real outputs of the guide, which are the rescaled radar image, the same image with the Lee filter and the three-class K-Means clustering that separates water, vegetation and buildings. (open the full-size image)
A real noisy patch, its averaged reference and the SwinIR output, next to the PSNR, SSIM and ENL of the autoencoder and of SwinIR, which ended below the input. (open the full-size image)

What it is about

Guides three and four of Computer Vision work with synthetic aperture radar. Radar sees through clouds, but its images carry speckle, a multiplicative noise that makes it hard to tell water, vegetation and buildings apart.

In guide three I prepared ten Sentinel-1 images of a port area. I rescaled them, filtered them with Lee, clustered them with K-Means and built a dataset of noisy and reference pairs. In guide four I compared networks for scene classification and tried two restoration models on those pairs.

What it does

  • Ten monthly Sentinel-1 images in VV polarization, from May 2025 to February 2026, rescaled and filtered with a 7 × 7 Lee filter.
  • K-Means with k = 3 and k = 4 on the filtered and unfiltered images, and a water versus land mask built from the clusters.
  • The ten dates, aligned and averaged, make a noise-free reference; 512 × 512 patches are cut from each pair for training.
  • In scene classification, ResNet50 with transfer learning reached 99.58% accuracy, while a sequential CNN trained from scratch overfit and collapsed to a single class.
  • An autoencoder and SwinIR to reduce speckle, which both lowered PSNR and SSIM against the input, and the report says so.

Decisions and their costs

  1. A reference made from the same images

    There is no noise-free version of a radar image to measure against.

  2. Publish that the restoration did not work

    The autoencoder lowered PSNR from 25.80 to 22.66 dB, and SwinIR, pretrained on optical photos, from 19.91 to 19.64 dB.

Repositories

Every public repository of the project, with what it holds.

  • 2026

    ComputerVisionIAEngineeringSebastian-s

    Computer vision in five guides, on convolutions and color, a vision model behind an API, Sentinel-1 radar clustering, deep learning and lettuce disease detection.

    Python, Jupyter
GitHub organization: EngineeringSebastian-s

Stack

Pick a technology to see the projects that use it.

Related