Skip to Main content Skip to Navigation
New interface

Deep learning for SAR imagery : from denoising to scene understanding

Abstract : Synthetic Aperture Radars (SARs) can collect data for Earth Observation purposes regardless of the daylight or cloud cover. Nowadays, thanks to the Copernicus program of the European Space Agency, a huge amount of SAR data is freely available. However, the exploitation of satellite SAR images is limited by the presence of strong fluctuations in the backscattered signal. Indeed, SAR images are corrupted by speckle, a phenomenon inherent to coherent imaging systems. In this Ph.D thesis, we aim to improve the interpretation of SAR images by resorting to speckle reduction techniques. Existing approaches are based on Goodman’s model, which describes the speckle component as a spatially uncorrelated multiplicative noise. In the computer vision field, denoising methods relying on Convolutional Neural Networks (deep learning approaches) have led to great improvements and provide nowadays state-of-the-art results. We propose to use deep learning-based denoising techniques to reduce speckle from SAR images (despeckling methods). At first, we study the adaptation of supervised techniques that minimize a certain distance between the estimation provided by the CNN and a reference image, also called “groundtruth”. We propose to create a dataset of reference images by averaging multi-temporal images acquired over the same area. Pairs of reference and corrupted images can be generated by synthetizing speckle following Goodman’s model. However, in real images the speckle component is spatially correlated which typically requires subsampling these images by a factor 2 to reduce the spatial correlations, which also degrades the spatial resolution. Given the limits of supervised approaches and inspired by noise2noise, a self-supervised denoising method, we propose to train our networks directly on actual SAR images. The principle of self-supervised denoising methods is the following: if a signal contains a deterministic component and a random component, then a network trained to predict a new signal realization from a first independent signal realization will only predict the deterministic component, i.e., the underlying scene, thereby suppressing the speckle. In the method we have developed, SAR2SAR, we leverage multi-temporal SAR series to obtain independent realizations of the same scene, under the hypothesis of temporally decorrelated speckle. Changes are compensated by devising an iterative training strategy. SAR2SAR is thus trained directly on images with spatially correlated speckle and can readily be applied on SAR images without subsampling, providing high-quality results. The training of SAR2SAR is quite heavy: it is articulated in several steps to compensate changes and a dataset comprising stacks of images must be built. With our approach “MERLIN”, we alleviate the training by proposing a single-image learning strategy. Indeed, in single-look-complex SAR images, real and imaginary parts are mutually independent and can benaturally exploited to train CNNs with self-supervision. We show the potential of this training framework for three imaging modalities, different in terms of spatial resolution, textures, and speckle spatial correlation. For the sake of open science, the code associated to each algorithm developed is made freely available.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Thursday, May 12, 2022 - 4:25:47 PM
Last modification on : Friday, May 13, 2022 - 9:20:01 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03666646, version 1



Emanuele Dalsasso. Deep learning for SAR imagery : from denoising to scene understanding. Artificial Intelligence [cs.AI]. Institut Polytechnique de Paris, 2022. English. ⟨NNT : 2022IPPAT008⟩. ⟨tel-03666646⟩



Record views


Files downloads