Satellite optical images, upon their on-ground receipt, offer a distorted view of the observed scene. Their restoration, including denoising, deblurring, and sometimes super-resolution, is required before their exploitation. Moreover, quantifying the uncertainties related to this restoration helps to reduce the risks of misinterpreting the image content. Deep learning methods are now state-of-the-art for satellite image restoration. Among them, direct inversion methods train a specific network for each sensor, and generally provide a point estimation of the restored image without the associated uncertainties. Alternatively, deep regularization (DR) methods learn a deep prior on target images before plugging it, as the regularization term, into a model-based optimization scheme. This allows for restoring images from several sensors with a single network and possibly for estimating associated uncertainties. In this paper, we introduce VBLE-xz, a DR method that solves the inverse problem in the latent space of a variational compressive autoencoder (CAE). We adapt the regularization strength by modulating the bitrate of the trained CAE with a training-free approach. Then, VBLE-xz estimates relevant uncertainties jointly in the latent and in the image spaces by sampling an explicit posterior estimated within variational inference. This enables fast posterior sampling, unlike state-of-the-art DR methods that use Markov chains or diffusion-based approaches. We conduct a comprehensive set of experiments on very high-resolution simulated and real Pléiades images, asserting the performance, robustness and scalability of the proposed method. They demonstrate that VBLE-xz represents a compelling alternative to direct inversion methods when uncertainty quantification is required.

Regularization of inverse problems is of paramount importance in computational imaging. The ability of neural networks to learn efficient image representations has been recently exploited to design powerful data-driven regularizers. While state-of-the-art plug-and-play (PnP) methods rely on an implicit regularization provided by neural denoisers, alternative Bayesian approaches consider Maximum A Posteriori (MAP) estimation in the latent space of a generative model, thus with an explicit regularization. However, state-of-the-art deep generative models require a huge amount of training data compared to denoisers. Besides, their complexity hampers the optimization involved in latent MAP derivation. In this work, we first propose to use compressive autoencoders instead. These networks, which can be seen as variational autoencoders with a flexible latent prior, are smaller and easier to train than state-of-the-art generative models. As a second contribution, we introduce the Variational Bayes Latent Estimation (VBLE) algorithm, which performs latent estimation within the framework of variational inference. Thanks to a simple yet efficient parameterization of the variational posterior, VBLE allows for fast and easy (approximate) posterior sampling. Experimental results on image datasets BSD and FFHQ demonstrate that VBLE reaches similar performance as state-of-the-art PnP methods, while being able to quantify uncertainties significantly faster than other existing posterior sampling techniques. The code associated to this paper is available in https://github.com/MaudBqrd/VBLE.

Earth observation through satellite images is crucial to help economic activities as well as to monitor the impact of human activities on ecosystems. Current satellite systems are subjected to strong computational complexity constraints. Thus, image compression is performed onboard with specifically tailored algorithms while image denoising is performed on the ground. In this letter, we intend to address satellite image compression and denoising with neural networks. The first proposed approach uses a single neural architecture for joint onboard compression and denoising. The second proposed approach sequentially uses a first neural architecture for onboard compression and a second one for on ground denoising. For both approaches, the onboard architectures are lightened as much as possible, following the procedure proposed by Alves de Oliveiraet al.(2021). The two approaches are shown to outperform the current satellite imaging system and their respective pros and cons are discussed.

The accurate on-orbit radiometric calibration of optical sensors has become a challenge for space agencies who combine their efforts through international working groups such as CEOS/WGCV or WMO/GSICS with the objective of ensuring the consistency of space measurements and to reach the absolute accuracy required by increasingly demanding scientific requirements. Different targets are generally used for calibration, depending on the specific features of the sensor or spacecraft: from on-board calibration systems to ground targets, they can all profitably be characterized and modeled. Thanks to their agility, some satellites have the capability to view extra-terrestrial targets such as stars. Although the spectral irradiance of some stars is known with very high accuracy, very little work has been done on their use as an absolute reference for the calibration of remote sensors. The reason is their low irradiance, requiring a small instantaneous field of view to observe them. Moreover, their use as a calibration reference also requires a good knowledge of the instrument's modulation transfer function (MTF). The MTF which drives the spatial quality is a major parameter for high-resolution (HR) optical remote sensors as it indicates how spatial frequencies are transmitted and weakened by the imaging system. On the other hand, stars provide ideal targets for MTF measurement as they can be assimilated to point sources. This paper proposes a method for simultaneously computing the absolute radiometric calibration of PLEIADES' HR optical sensors and their MTF, using stars. The radiometric model is solved in Fourier space for point sources whose irradiance can be easily controlled. Results are provided for both PLEIADES satellites and compared to the official radiometric calibration and MTF. The quality of the results and the main error contributors are discussed.

In the context of the Spot World Heritage (SWH) initiative, the long term archive of Spot 1 to 5 satellite data will be reprocessed. This initiative gives the opportunity to increase Spot data quality and production efficiency with the introduction of new state-of-the-art processing methods. In this context, we consider the Spot 5 Supermode processing chain and the introduction of a new denoising method. Our objective is to propose the best denoising method in terms of efficiency and accuracy for the Spot 5 Supermode data production. We report two experimentations that lead to the choice of the Non Local Bayes (NLB) denoising method; it is fast, accurate and remains stable in terms of efficiency and quality for different landscapes and image sizes. We also introduce two optimized versions of the NLB algorithm that increase the computation efficiency by a factor up to 4.

Since SPOT1, the French national space center (CNES) has worked on improving the geometry of Earth observation spacecrafts. The accuracy of sensor calibration is one of the main key points for any Earth observation application such as orthorectification, DEM generation or surface change detection [7]. Two families of methods have been developed by CNES for twenty years: absolute methods and relative methods. These methods are used to characterize a pushbroom acquisition along the detector line and the time line. By this way, the viewing directions are measured and the residual of the spacecraft's attitude angles (not restituted by the AOCS) are estimated. This information can complete the geometric model of all the scenes acquired by the instrument and is used in all geometric applications. We will first consider the absolute methods and then the relative methods.

The new French high resolution earth observing satellite PLEIADES-HR will be launched in 2011. A specific design and new technologies have been embarked to provide great agility. These capabilities offer new methods to perform the image quality activities during the commissioning period. Some of them depend on dedicated guidance of the satellite platform and specific targets on the Earth or beyond. For instance, one concerns the so-called AMETHIST method to compute the normalization coefficients of the radiometric model. Another one is based on a controlled slow-motion to get a reference image line whose deviation along the image will give information on radiometric noise or attitude perturbations. Another uses the stars to measure the line-of-sight dynamic stability or the instrument refocusing needs. These originals methods and some more are briefly presented in this paper after an introduction of PLEIADES-HR capacities.

3D information from DTM or DEM is widely used to enhance understanding of satellite images. 3D navigation tools usually include DTM data. Large stereoscopic angles is the common rule since DTM extraction process will then be less sensitive to geometric model accuracy. However, low stereoscopic angle offer valuable evens such as high similarity between the two images of the stereoscopic couple to be correlated. Further more, some satellite such as SPOT5, Formosat, Venus (to be launched) offer native stereoscopy with low B/H value, with CCD arrays staggered in the same focal plane. This paper focuses on the case of SPOT5, where one takes benefit from the low stereoscopic angle between Panchromatic and Multispectral images in order to extract Terrain Model and simultaneously refine the geometric model.

Since early 1999, the CNES has worked on defining and equipping supersites for its needs in image quality assessment. Various criteria guided the definition of this supersite involving the characteristics of high resolution earth observation satellites and the image quality assessment methods. The main goal of this work is to have high resolution geometric references in order to calibrate any high resolution spatial sensor. The accuracy of sensor calibration that can be obtained using image matching techniques depends both on planimetric and altimetric accuracy of the reference. In order to produce this kind of references, very high resolution aerial images are used (10 cm ground sampling distance). Two very high resolution cameras fly over several sites in order to produce stereoscopic images with narrow angular fields of view. This original acquisition system allows us to produce a high resolution 3D reference with similar images containing no shadows effects. Moreover, thanks to this system, the overlap between a couple of images is nearly 100% if the stereoscopic angle is low. One of the key factors of the quality of the 3D reconstruction is the value of the stereoscopic angle used. This paper shows that low stereoscopic angles are sufficient to build the elevation model including superstructures. This characteristic is the determining factor to make spatial native stereoscopy. The essential condition to compute an high accurate elevation model with the aerial data is to control the geometrical model of all the images. A lack of knowledge of the geometrical model could be interpreted as a default of altitude of the scene (or a part of the scene). Therefore to build the digital surface model (DSM), the planimetry is first controlled through particularly accurate aerotriangulation based on lot of tie points and also, if available, based on ground control points measured with high accuracy through a GPS measurement campaign. The accurate geometry of all the images is computed with a complete aerotriangulation. Once the geometrical models are accurately known, a high resolution DSM can be created using a dense correlation process between images couples with a low stereoscopic angle. The aerotriangulation and the DSM allow us to have a high resolution geometric reference that ensures both planimetric and altimetric accuracy which are key factors for the geometric quality of a spatial sensor calibration. This kind of reference will be used to calibrate a spatial sensor like Pleiades.

Following SPOT satellite family, PLEIADES-HR is a high resolution optical earth observation system developed by CNES and consisting in two satellites with a first launch by the end of 2009. Pleiades-HR will deliver both 70 cm Panchromatic images and 2.8 m multispectral (blue-green-red-near infrared) images with a 21km swath. As for SPOT images, Pleiades-HR ground segment includes a systematic and automatic cloud notation in order to feed the catalogue with a binary cloud mask and a confidence level notation used to monitor a manual re- notation in case of low confidence level. The cloud detection is operated from preview images generated through a fast decompression step of the low frequency wavelet subband : since Pleiades-HR compression is based upon a 3 levels wavelet transform, extracting the low frequency subband from the panchromatic and the multispectral images leads respectively to 5.6 m panchromatic and 22.4 multispectral images. In order to significantly improve the SPOT cloud notation process and get rid of frequent manual renotation, an innovative cloud detection technique has been studied and finally specified. It takes benefit from two kinds of information issued from the panchromatic and multispectral previews : multispectral reflectances and geometric parallaxes due to a panchromatic/multispectral stereoscopic angle that makes panchromatic/multispectral misregistration slightly sensitive to cloud altitudes and displacement velocity. These radiometric and geometric criteria are used as inputs of a Support Vector Machine (SVM) algorithm. SVM's are widely used for detection purposes and are based upon a training set of pixels with known criteria and classification (cloudy or cloud/free). Basically, the SVM technique aims to geometrically separate the training set represented in a Rnspace, with n standing for the number of radiometric and geometric criteria taken into account for classification, using an hyperplane or some more complex surface if necessary. SVM training algorithm finds out the best frontier in order to maximize the margin, defined as a symmetric zone centered on the frontier with no training points included, and to minimize the number of wrong classification occurrences. In order to reach that goal, SVM training algorithm usually implements a Lagrangian minimization technique. Among SVM evens, one may quote the reduced complexity for the detection step : using an hyperplane to separate the two classes means that the detection process resumes to a scalar product in the Rnspace for each pixel. Apart from the computation of radiometric and geometric criteria associated to each pixel to be classified, the complexity mainly lies in the training stage, executed only once for instance during the inflight commissioning period. Another advantage is the ability to generate a confidence mark for each pixel classification based upon the distance measured in the Rnspace between the frontier and the point representative of the pixel to be classified : the general rule is that a large distance means a high confidence mark. The SVM technique has been implemented and assessed using a Quickbird Panchromatic/Multispectral data base, with various cloud coverages. Quickbird data are representative of future Pleiades images both in terms of resolution, spectral bands and even panchromatic/multispectral stereoscopic angle : it was thus possible to simulate quite representative panchromatic/multispectral quicklooks. The study also implied the generation of reference cloud masks that were produced through visual detection. A whole sensitivity analysis including the radiometric/geometric criteria taken into account and SVM parameters was conducted. Results were classically analyzed in terms of good detection and false alarm performances. The paper details the whole study of this SVM cloud detection technique within PLEIADES-HR framework.

Pleiades-HR is a high-resolution optical Earth observation system developed by CNES, for civilian and military users. The launch of the first Pleiades satellite is scheduled in 2008, the second, 18 months later. The Pleiades-HR program is the French part of the French-Italian ORFEO program which also comprises COSMO-SkyMed, an Italian high-resolution radar system. The Pleiades-HR products will be distributed by Spotlmage and will complete the existing SPOT product line. In such a multi-sensors and multi-users context, CNES has decided to design the new Pleiades-HR ground segment pointing the stress on systems interoperability in order to ease the data access from the end user point of view. After a synthetic presentation of the Pleiades HR ground segment and image products, this paper focuses on PLEIADES HR image product definition and cataloguing services, two major points for the image ground segment interoperability. Among all possible solutions, the image product and catalog services have been defined taking into account the PLEIADES-HR system characteristics, its performance requirements and the international standards (ISO-TC211, OGC, image standards, JPEG2000 ...) use and evolution capabilities. The results presented in the paper are also part of a European initiative of ground segment harmonisation studied in the frame of the HMA project with other European Space Agencies.

Remote sensing optical images are often cloudy and then partially unusable. Thus, cloud detection can optimize the image acquisition loop and the end-user image selection. The current pre-processing of SPOT images includes an automatic cloud and snow detection algorithm based on neural networks and fuzzy logic, which globally provides correct cloud masks but with a perfectible confidence. This process must be improved in order to avoid systematic manual re-notation. For Ple/spl acute/iades high resolution (HR) satellite, CNES experiments an innovative cloud detection method based on correlation and stereoscopic effect (B/H ratio) between images. Thanks to the quasi-simultaneous acquisition of the five spectral bands (panchromatic, red, blue, green and near infrared), this new method can be systematically applied, in addition to a radiometric one. This paper focuses on the new detection algorithm for high resolution images, which includes six main steps: (1) generation of 20 m resolution preview images in both multispectral and panchromatic bands, (2) estimation of the misregistration between panchromatic and multispectral previews using a geometric model, a global DTM and an image matching algorithm, (3) computation of the residual parallax consisting in the difference between prediction model and image matching, (4) cloud detection through high parallax value thresholding (5) radiometric analysis for snow and low altitude clouds detection, (6) masks fusion and confidence status computation. This method, still under assessment over various SPOT5 and Quickbird images, seems to be very promising. The results, presented in the paper, show that the method is efficient even for Quickbird images, in spite of the very low B/H ratio value (0.002) close to Ple/spl acute/iades one.

Following SPOT5 launch, Spotlmage and French National Geographic Institute (IGN) have decided to design a worldwide accurate database called Reference3D/sup TM/ using data from the High Resolution Stereoscopic SPOT5 instrument (HRS). Spotlmage and IGN have also decided to develop and commercialize a system called ANDORRE to produce orthorectifled images thanks to Reference3D/sup TM/ data. ANDORRE has been designed to take advantage of Reference3D/sup TM/ planimetric and altimetric accuracy to automatically register and rectify any image from SPOT satellites. In this framework, CNES is acting as prime contractor to realize and industrialize the algorithms following a preliminary study undertaken by IGN. This paper focuses on the validation of the geometric performances of the algorithms. A special attention is also paid to the processing time in order to achieve the rectification of a 24000/spl times/24000 image in less than one hour.

In order to increase SPOT5 panchromatic resolution, CNES adopted in 1995 a quincunx sampling mode named THR, which is a French acronym for "very high resolution". Such a sampling is produced by two CCD linear arrays shifted in the focal plane and has been shown to meet Shannon requirement at first order. That means a good fit between the sampling and the instrument modulation transfer function (MTF) and thus an optimization of the whole acquisition system. SPOT5 THR is a 2.5 m image obtained through a rather complex processing of the two 5 meter shifted images delivered by the double CCD linear array, consisting in quincunx interpolation, deconvolution and denoising. Quincunx interpolation computes radiometric information over a 2.5m grid while only half the information may be retrieved from the two 5m images, deconvolution compensates for low MTF values for high spatial frequencies and denoising reduces the noise level enhancement due to deconvolution. Apart from the high resolution panchromatic mode, SPOT5 also delivers classical 10m multispectral images covering the green, red and near infrared spectral domains that are almost simultaneously acquired. Merging the 2.5m THR with the 10m multispectral yields a 2.5m multispectral product which proved to be of utmost interest for thematic applications. Gathering the main results from the SPOT5 in orbit commissioning period, this paper gives an overview of THR and fusion processing, points out onboard and onground sensitive parameters and finally presents the THR and THX performances.

This paper considers the estimation scenario where an original image has undergone blurring, noise corruption and compression. The authors consider these unwanted effects jointly in the estimation formalism they propose. The contribution of the paper aims to address the joint coding/decoding formulation, involving a priori assumptions on the solution and knowledge of the imaging systems to account for effects due to acquisition noise and compression noise. Most of the authors' construction is based on well-known techniques drawn from a variety of areas in modern signal processing, including optimization theory, wavelet decomposition, optimal bit allocation, inverse problem and bounded noise assumption.

We propose a new multiresolution variational joint denoising/deblurring approach, involving a priori assumptions on the solution and knowledge of the imaging systems to account for effects due to acquisition noise (edge preservation, degradation noise modeling, bounded noise assumption and spectral control of noise level-whiteness and stationarity). The techniques used are drawn from a variety of areas of modern signal processing, including optimization theory, inverse problem, wavelet packets decomposition and bounded noise assumption.