Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A speckle filter for Sentinel-1 SAR Ground Range Detected data based on Residual Convolutional Neural Networks

Alessandro Sebastianelli, Maria Pia Del Rosso, Silvia Liberata Ullo, Paolo Gamba

TL;DR

This work aims at presenting a novel method for filtering speckle noise from Sentinel-1 ground range detected data by applying deep learning algorithms, based on convolutional neural networks, proving the effectiveness of the proposed architecture.

Abstract

In recent years, machine learning (ML) algorithms have become widespread in all the fields of remote sensing (RS) and earth observation (EO). This has allowed the rapid development of new procedures to solve problems affecting these sectors. In this context, this work aims at presenting a novel method for filtering speckle noise from Sentinel-1 ground range detected (GRD) data by applying deep learning (DL) algorithms, based on convolutional neural networks (CNNs). The paper provides an easy yet very effective approach to extract the large amount of training data needed for DL approaches in this challenging case. The experimental results on simulated speckled images and an actual SAR dataset show a clear improvement with respect to the state of the art in terms of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), equivalent number of looks (ENL), proving the effectiveness of the proposed architecture.

A speckle filter for Sentinel-1 SAR Ground Range Detected data based on Residual Convolutional Neural Networks

TL;DR

This work aims at presenting a novel method for filtering speckle noise from Sentinel-1 ground range detected data by applying deep learning algorithms, based on convolutional neural networks, proving the effectiveness of the proposed architecture.

Abstract

In recent years, machine learning (ML) algorithms have become widespread in all the fields of remote sensing (RS) and earth observation (EO). This has allowed the rapid development of new procedures to solve problems affecting these sectors. In this context, this work aims at presenting a novel method for filtering speckle noise from Sentinel-1 ground range detected (GRD) data by applying deep learning (DL) algorithms, based on convolutional neural networks (CNNs). The paper provides an easy yet very effective approach to extract the large amount of training data needed for DL approaches in this challenging case. The experimental results on simulated speckled images and an actual SAR dataset show a clear improvement with respect to the state of the art in terms of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), equivalent number of looks (ENL), proving the effectiveness of the proposed architecture.

Paper Structure

This paper contains 14 sections, 19 equations, 10 figures, 7 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Data and Methods
  4. Dataset creation for the proposed CNN
  5. Proposed approach
  6. Results and Discussion
  7. Results on COCO dataset
  8. Results and comparisons on the reference dataset
  9. Results and comparisons on a real GRD image
  10. Further analysis
  11. Edges Preservation
  12. Preservation of the statistical characteristics
  13. Conclusions
  14. Speckle Model

Figures (10)

  • Figure 1: Comparison between Sentinel-1 (on the left) and Sentinel-2 (on the right) acquisitions on 2022-03-17 and 2022-04-30 respectively, focusing on Rome (Italy): a Single Look Complex (SLC) Interferometric Wide Swath (IW) mode SAR image in VV polarization versus a Level2 (L2)-A optical image in true colors (composition of RGB channels).
  • Figure 2: Dataset creation pipeline. On top: step 1 - Download of SAR time-series from GEE and application of the temporal average to get the ground truth. On bottom: step 2 - generation of speckle noise with Gamma distribution to be multiplied with the ground truth to get the noisy input (step 3).
  • Figure 3: Proposed model for speckle filtering.
  • Figure 4: Graphical results of the proposed model tested by adding Gaussian noise to the gray-scale version of the COCO dataset. Several simulation were made by varying the standard deviation of the Gaussian noise while keeping the mean equal to zero.
  • Figure 5: Graphical results of the proposed model tested by adding the simulated Speckle noise to the gray-scale version of the COCO dataset. Several simulation were made by varying the number of Looks.
  • ...and 5 more figures