Deep learning for remote sensing images and their interpretation

Overview & Synopsis

Synthetic Aperture Radar (SAR) images are not affected by the presence of clouds or variations of sunlight. They provide very useful information for Earth observation (chapter 1).They are impacted by strong fluctuations called "speckle" which make their interpretation difficult. The speckle is a phenomenon intrinsic to the coherent illumination of the scene by the radar, meaning that speckle-free images can not be captured and used as reference to train models.The properties of speckle are different from that of the traditional additive white Gaussian noise used to model corruptions in optical images, and proper despeckling algorithms are needed. Most of them rely on statistics derived from the Goodman's model (chapter 2). Recently, deep learning based methods have been very successful at despeckling a single SAR image. This work focuses on improving the despeckling performance by jointly processing several input images to exploit the common information while still preventing the propagation of differences from one image to another (chapter 3).The despeckling of Sentinel-1 GRDM Extra Wide images of sea ice is studied in Chapter 4 for sea ice studies. The ice is shifting quickly on the sea and multi-temporal stacks of a specific area can not be used for despeckling purposes due to structural changes. In the images, thermal noise can not be neglected because the reflectivity values of water and ice are very low and close to the thermal noise floor. We propose a dual-polarimetric despeckling framework where HH and HV polarimetric channels are used as input and are jointly despeckled in a single pass. The network is trained in a self-supervised way inspired by the existing SAR2SAR framework and takes corrected images where the thermal noise floor level has been removed as input. Our approach shows a clear improvement over existing image restoration techniques on Sentinel-1 images of the Artic.Despeckling can be improved by combining measurements pertaining to common information within the temporal stack while ignoring data impacted by temporal changes.First, multi-temporal despeckling methods using temporal averaging and the computation of a super-image (i.e. despeckled temporal mean image) are introduced at the beginning of Chapter 5. A generative model is then proposed to explicit the statistics of SAR multi-temporal stacks and account for the spatial and temporal correlations of speckle. A multi-temporal extension of the existing MERLIN framework is derived from this model. The network is fed with additional images of the same area acquired at different dates. It is trained in an unsupervised way inspired by the Noise2Noise framework: the real part (or the imaginary part) of the image and additional dates are fed to the network and the imaginary part (or the real part) is used as a target. Adding more images continuously improves the despeckling performance, but with diminishing gains.A temporal whitening is proposed to prevent the drop of performance of the network when the input channels are temporally correlated.Despeckling methods are hard to evaluate because of the lack of ground truth images. Chapter 6 focuses on uncertainty quantification for despeckling using deep learning.First, works are presented to combine despeckling and estimation of the uncertainty map during the training. Starting from a framework where only one value is predicted for each pixel, we aim at predicting a distribution for each pixel. Parameters of uniform and inverse gamma distributions are estimated. The sharper the distribution, the more certain the network is of its prediction. We discuss the difficulty of estimating uncertainties in a self-supervised learning framework where the noise level is high and the limits faced by our formulations.Working with the MERLIN framework, an estimation of an uncertainty map is proposed based on the expected difference map between predictions from the real and imaginary parts.