Search

Hyperspectral images provide detailed spectral information through hundreds of (narrow) spectral channels (also known as dimensionality or bands) with continuous spectral information that can accurately classify diverse materials of interest. The increased dimensionality of such data makes it possible to significantly improve data information content but provides a challenge to the conventional techniques (the so-called curse of dimensionality) for accurate analysis of hyperspectral images. Feature extraction, as a vibrant field of research in the hyperspectral community, evolved through decades of research to address this issue and extract informative features suitable for data representation and classification. The advances in feature extraction have been inspired by two fields of research, including the popularization of image and signal processing as well as machine (deep) learning, leading to two types of feature extraction approaches named shallow and deep techniques. This article outlines the advances in feature extraction approaches for hyperspectral imagery by providing a technical overview of the state-of-the-art techniques, providing useful entry points for researchers at different levels, including students, researchers, and senior researchers, willing to explore novel investigations on this challenging topic. In more detail, this paper provides a bird's eye view over shallow (both supervised and unsupervised) and deep feature extraction approaches specifically dedicated to the topic of hyperspectral feature extraction and its application on hyperspectral image classification. Additionally, this paper compares 15 advanced techniques with an emphasis on their methodological foundations in terms of classification accuracies. Furthermore, the codes and libraries are shared at https://github.com/BehnoodRasti/HyFTech-Hyperspectral-Shallow-Deep-Feature-Extraction-Toolbox.

This paper describes a novel framework for compressive sampling of multichannel signals that are highly correlated across the channels. In this work, we assume few number of independent sources are generating the multichannel observations based on a linear mixture model. Moreover, sources are assumed to have sparse/compressible representations in some orthonormal basis. The main contribution of this paper lies in rephrasing the compressed sampling of multichannel data as the compressive source separation problem by knowing the mixture parameters. A number of simulations measure the performance of our recovery algorithm. Comparing to the classical CS scheme -which recovers data of all channels separately- ours indicates a significant reduction in both, the number of measurements to be sent and the complexity of the decoding algorithm (i.e., scaling by the number of sources rather than the number of channels). We demonstrate an application of our scheme in acquisition of the Hyperspectral images. Our algorithm proposes an accurate, low cost recovery from small number of the transmitted measurements that are taken by sensors with very low spatial resolution.

Hyperspectral images (HI) and Light Detection and Ranging (LiDAR) provide high resolution radiometric and geometric information for monitoring forests at individual tree crown (ITC) level. It has many important applications for sustainable forest management, biodiversity assessment and healthy ecosystem preservation. However, the integration of different remote sensing modalities is a challenging task for tree species classification due to different artifacts such as the lighting variability, the topographic effects and the atmospheric conditions of the data acquisition. The characterization of ITC can benefit from the extraction and selection of robust feature descriptors that solve these issues. This paper aims to investigate the integration of feature descriptors from HI and LiDAR by using the intra-set and inter-set feature importance for the semantic segmentation of forest tree species. A fusion methodology is proposed between high-density LiDAR data – (20 pulses m−2) and VNIR HI – (160 bands and 0.80 m spatial resolution) acquired on French temperate forests along an altitude gradient. The proposed scheme has three inputs: the field inventory information, the HI and the LiDAR data. Our approach can be described in nine stages: polygon projection, non-overlapping pixel selection, vegetation and shadow removal, LiDAR feature extraction, height mask, robust PCA (rPCA), feature reduction and classification. The overall accuracy of tree species classification at pixel-level was 68.9% by using random forest (RF) classifier. Our approach showed that 74.0% of trees were correctly assigned overall, by having conifer species such as Norway Spruce (Picea abies) with a producer’s accuracy of 97.4%.

International audience; The advantage of hyperspectral image is the possibility to access the spectral signature for each pixel having the capability to see the unseen. The classifiers used for hyperspectral images can be deeply studied and analyzed for obtaining an accurate classification to extract the features from the images. In this paper we aim to analyze the hyperspectral images by trying to identify the nature of each element in the field, which is represented by a pixel, using the techniques of supervised learning.

International audience; This paper introduces a matrix cofactorization approach to perform spectral unmixing and classification jointly. After formulating the unmixing and classification tasks as matrix factorization problems, a link is introduced between the two coding matrices, namely the abundance matrix and the feature matrix. This coupling term can be interpreted as a clustering term where the abundance vectors are clustered and the resulting attribution vectors are then used as feature vectors. The overall non-smooth, non-convex optimization problem is solved using a proximal alternating linearized minimization algorithm (PALM) ensuring convergence to a critical point. The quality of the obtained results is finally assessed by comparison to other conventional algorithms on semi-synthetic yet realistic dataset.

Short range hyperspectral imagery is a promising tool for phenotyping and vegetation survey. When associated with partial least square regression (PLS-R), it allows high spatial resolution mapping of the plant chemical content at the canopy scale. However, several optical phenomena have to be taken into account when applying this approach to vegetation scenes in natural conditions. For instance, additive and multiplicative factors due respectively to specular reflection and leaf inclination can be overcome by spectral preprocessing. But the most challenging phenomenon is multiple scattering. It appears when a leaf is partly lightened by the reflected or transmitted light from surrounding leaves, resulting in strong non linear effects in its apparent reflectance spectrum. Though this effect can be taken into account in some remote sensing models at the canopy scale, no study has been proposed until now concerning its impact on spectral prediction of vegetation chemical content by short range imagery.The objective of this project, associated with a PhD work, was to analyze these effects in the context of hyperspectral imagery for vegetation phenotyping purpose, and to propose spectral processing methods to overcome them.The methodological development has been based on simulation tools included in the open source platform OpenAlea (http://openalea.gforge.inria.fr/dokuwiki/doku.php). A typical wheat canopy scene has been modelled using Adel-Wheat and combined with the light propagation model Caribu. The proposed tool simulates the apparent reflectance of every visible leaf in the canopy for a given actual reflectance and transmittance, allowing to synthetize realistic hyperspectral images.This simulation approach has allowed us, in a first step, to analyze the distribution of deviations due to multiple scattering in the spectral space, and then to infer a correction method in the frame of PLS regression. This method relies on the building of two subspaces EW and EB respectively generated by the analytic formulation of multiple scattering and by the variable of interest. It allows us to define a projection operation on EB subspace along EW direction (oblique projection), in order to remove multiple scattering effects while preserving useful information. This projection operation is then applied on every spectra during learning phase and using phase of the PLS model.The method has first been developed and tuned using simulated data, in the frame of leaf nitrogen content (LNC) prediction of wheat leaves. For this purpose, reflectance spectra (450-1100 nm) of 57 wheat leaves have been collected using a ASD filed spectrometer (FieldSpec®, Analytical Spectral Devices, Inc., Boulder, Colorado, USA), while their LNC was measured through reference chemical analyses. Regression models with and without oblique projection have then been built from the ASD spectra and applied to simulated data. The model with oblique projection provided excellent results (R² = 0.931; RMSEP = 0.29% DM), compared to the classical one (R² = 0.915; RMSEP = 0.42% DM).The same method has then been applied in real conditions on wheat pot plants and field plants. For this purpose, some leaves have been collected and laid on a black paper background to be imaged, in order to build PLS models that have then been applied on in-situ plants. These experimentations have confirmed that the classical PLS-R induces a strong overestimation of LNC on leaves surrounded by other leaves, and that oblique projection corrects this overestimation (same prediction on surrounded then isolated leaf).; L'imagerie hyperspectrale de proximité est un outil prometteur pour le phénotypage ou la surveillance de la végétation. En association avec la régression des moindres carrés partiels ou PLS-R, elle permet de construire des cartographies de haute résolution spatiale du contenu chimique à l’échelle de la canopée. Cependant, plusieurs phénomènes optiques doivent être pris en compte lors de l'application de cette approche aux scènes de végétation dans des conditions naturelles. Notamment, les facteurs additifs et multiplicatifs liés respectivement à la réflexion spéculaire et à l'inclinaison des feuilles qui peuvent être surmontés par prétraitement. Mais le phénomène qui pose le plus de défis est la réflexion multiple. Il se produit lorsqu'une feuille est éclairée en partie par la lumière directe, et en partie par la réflexion ou la transmission de la lumière des feuilles voisines, induisant de forts effets non linéaires sur son spectre de réflectance. Bien que cet effet puisse être pris en compte dans certains modèles de télédétection à l’échelle de la canopée, aucune étude n’a été proposée à ce jour sur la façon dont un tel phénomène affecte les évaluations spectrales de la biochimie végétale par imagerie de proximité. L'objectif de la présente étude était d'analyser ces effets dans le contexte de l'imagerie hyperspectrale à des fins de phénotypage végétal et de proposer des méthodes chimiométriques pour les surmonter. Le développement méthodologique a été basé sur des outils de simulation inclus dans la plate-forme open source OpenAlea (http://openalea.gforge.inria.fr/dokuwiki/doku.php). Une scène typique de canopée de blé a été modélisée à l'aide du modèle Adel-Wheat et combinée au modèle de propagation de la lumière Caribu. L'outil proposé simule la réflectance apparente de chaque feuille visible dans la canopée pour une réflectance et une transmittance réelles données, permettant de synthétiser des images hyperspectrales réalistes. Cette approche par simulation nous a permis, dans un premier temps, d’analyser la distribution dans l’espace spectral des perturbations engendrées par les réflexions multiples, puis d’en déduire une méthode de correction applicable dans le cas d’une régression PLS. La méthode est basée sur la construction de deux sous-espaces W et B générés respectivement par la formulation analytique des réflexions multiples et la variable d'intérêt. Ceci nous permet alors de définir une matrice de projection sur B selon la direction W (projection oblique), qui permet de supprimer l’effet des réflexions multiples tout en conservant l’information utile. Il suffit ensuite d’appliquer cette projection à chaque spectre lors de l’apprentissage et de la mise en œuvre du modèle PLS. La méthode a d’abord été développée et paramétrée sur les données simulées, dans le contexte de l’évaluation de la teneur en azote (LNC) de feuilles de blé. Pour cela, les spectres de réflectance (450-1100 nm) de 57 feuilles de blé ont été collectés à l'aide d'un spectromètre ASD (FieldSpec®, Analytical Spectral Devices, Inc., Boulder, Colorado, USA), tandis que leur LNC a été mesuré à l'aide d'analyses chimiques. Des modèles de régression avec et sans projection oblique ont alors été construits à partir des spectres ASD et appliqués sur l’ensemble des données simulées. Le modèle avec projection oblique a donné d’excellents résultats (R² = 0.931; RMSEP = 0.29% DM) en comparaison du modèle classique (R² = 0.915; RMSEP = 0.42% DM).La même méthode a ensuite été appliquée en conditions réelles, sur des feuilles de blé cultivées en pot et au champ. Pour cela, des feuilles ont été collectées et imagées à plat sur fond noir pour la construction des modèles PLS, qui ont ensuite été appliqués aux plantes sur pied. Ces expérimentations ont confirmé d’une part que la PLS-R classique entraînait une forte surestimation du LNC sur les feuilles entourées d’autres feuilles, d’autre part que la projection oblique évitait cette surestimation.

International audience; Lossy compression of images is used in different applications including remote sensing. A problem is how to carry out such lossy compression automatically with providing appropriate quality. This paper deals with considering peculiarities of using JPEG2000-based coders for compressing hyperspectral data. Two approaches component-wise and 3D are analyzed. It is demonstrated that 3D compression provides better results but both approaches are characterized by the same drawback it is problematic to provide distortions less or comparable with noise in hyperspectral images.

The circum-galactic medium consists in gas orbiting around galaxies, whose faintness prevents any complete and easy detection. A powerful tool to detect such pattern can be found in using hyperspectral imaging. Nevertheless, detection in hyperspectral datacubes faces various problems, including well-fitted signal and noise descriptions to ensure further discrimination. A specificity of astronomical images resides in dealing with faint and very noisy signals. In this paper, we introduce a new constrained generalized likelihood ratio test adapted to the problem and a compound test to exploit most of the available information. We also investigate the use of both spatial information and multiple observations on a single scene, to enhance robustness. Numerical experiments on synthetic data are performed to quantify the gain of the different approaches. Finally, results on real hyperspectral astronomical data are presented, which may map for the first time observation of the circum-galactic medium around faint and distant galaxies. HighlightsWe aim at detecting halos near faint galaxies in astronomical hyperspectral images.We introduce constrained GLR tests to target the fainter spectra.Extensions consider spatial enhancement and the use of several observations.Experiments on synthetic data enables the thorough evaluation of the methods.Results on real images could provide the first statistical mapping of the halos.

In hyperspectral images, some spectral bands suffer from low signal-to-noise ratio due to noisy acquisition and atmospheric effects, thus requiring robust techniques for the unmixing problem. This paper presents a robust supervised spectral unmixing approach for hyperspectral images. The robustness is achieved by writing the unmixing problem as the maximization of the correntropy criterion subject to the most commonly used constraints. Two unmixing problems are derived: the first problem considers the fully-constrained unmixing, with both the non-negativity and sum-to-one constraints, while the second one deals with the non-negativity and the sparsity-promoting of the abundances. The corresponding optimization problems are solved efficiently using an alternating direction method of multipliers (ADMM) approach. Experiments on synthetic and real hyperspectral images validate the performance of the proposed algorithms for different scenarios, demonstrating that the correntropy-based unmixing is robust to outlier bands., Comment: 23 pages

This work was supported in part by the National Science Fund for Excellent Young Scholars under Grant 61522107 and in part by the Natural Science Foundation of China under Grant 6137118; International audience; In this paper, we propose a novel multiple kernel learning (MKL) framework to incorporate both spectral and spatial features for hyperspectral image classification, which is called multiple-structure-element nonlinear MKL (MultiSE-NMKL). In the proposed framework, multiple structure elements (MultiSEs) are employed to generate extended morphological profiles (EMPs) to present spatial-spectral information. In order to better mine interscale and interstructure similarity among EMPs, a nonlinear MKL (NMKL) is introduced to learn an optimal combined kernel from the predefined linear base kernels. We integrate this NMKL with support vector machines (SVMs) and reduce the min-max problem to a simple minimization problem. The optimal weight for each kernel matrix is then solved by a projection-based gradient descent algorithm. The advantages of using nonlinear combination of base kernels and multiSE-based EMP are that similarity information generated from the nonlinear interaction of different kernels is fully exploited, and the discriminability of the classes of interest is deeply enhanced. Experiments are conducted on three real hyperspectral data sets. The experimental results show that the proposed method achieves better performance for hyperspectral image classification, compared with several state-of-the-art algorithms. The MultiSE EMPs can provide much higher classification accuracy than using a single-SE EMP.

International audience; Morphological attribute profiles are among the most prominent spectral-spatial pixel description methods. They are efficient, effective and highly customizable multi-scale tools based on hierarchical representations of a scalar input image. Their application to multivariate images in general, and hyperspectral images in particular, has been so far conducted using the marginal strategy, i.e. by processing each image band (eventually obtained through a dimension reduction technique) independently. In this paper, we investigate the alternative vector strategy, which consists in processing the available image bands simultaneously. The vector strategy is based on a vector ordering relation that leads to the computation of a single max-and min-tree per hyperspectral dataset, from which attribute profiles can then be computed as usual. We explore known vector ordering relations for constructing such max-trees and subsequently vector attribute profiles, and introduce a combination of marginal and vector strategies. We provide an experimental comparison of these approaches in the context of hyperspectral classification with common datasets, where the proposed approach outperforms the widely used marginal strategy.

International audience; This paper studies a fast multi-band image fusion algorithm for high-spatial low-spectral resolution and low-spatial high-spectral resolution images. The popular forward model and the conventional Gaussian prior are used to form the posterior of the target image. Maximizing the posterior leads to solving a matrix Sylvester equation. By exploiting the properties of the circulant and decimation matrices associated with the fusion problem, a closed-form solution for the corresponding Sylvester equation is obtained, avoiding any iterative update step. Simulation results show that the proposed approach using this closed-form solution achieves the same performance as existing algorithms with the advantage of significantly decreasing the computational complexity of these algorithms.

International audience; In several remote sensing applications, detecting exceptional/irregular regions (i.e, pixels) with respect to the whole dataset homogeneity is regarded as a very interested issue. Currently, this is limited to the pre-processing step aiming to eliminate the cloud or noisy pixels. In this paper, we propose to extend the coverage area and to tackle this issue by regarding the irregular/exceptional pixels as outliers. The main purpose is the adaptation of the class outlier mining concept in order to find abnormal and irregular pixels in hyperspectral images. This should be done taking into account the class labels and the relative uncertainty of collected data. To reach this goal, the Class Outliers: DistanceBased (CODB) algorithm is enhanced to take into account the multivariate high-dimensional data and the concomitant partially available knowledge of our data. This is mainly done by using belief theory and a learnable task-specific similarity measure. To validate our approach, we apply it for vegetation inspection and normality monitoring. For experimental purposes, the Airborne Prism Experiment (APEX) data, set acquired during an APEX flight campaign in June 2011, was used. Moreover, a collection of simulated hyperspectral images and spectral indices, providing a quantitative indicator of vegetation health, were generated for this purpose. The encouraging obtained results can be used to monitor areas where vegetation may be stressed, as a proxy to detect potential drought.

This paper deals with a problem of dimensionality reduction for hyperspectral images using principal component analysis. Hyperspectral image reduction is improved by adding structural/spatial information to the spectral information, by means of mathematical morphology tools. It can be then useful for instance in supervised classification. The key element of the approach is the computation of a covariance matrix which integrates simultaneously both spatial and spectral information.

International audience; This paper investigates the use of Gaussian processes to detect non-linearly mixed pixels in hyperspectral images. The proposed technique is independent of nonlinear mixing mechanism, and therefore is not restricted to any prescribed nonlinear mixing model. The observed reflectances are estimated using both the least squares method and a Gaussian process. The fitting errors of the two approaches are combined in a test statistics for which it is possible to estimate a detection threshold given a required probability of false alarm. The proposed detector is compared to a robust nonlinearity detector recently proposed using synthetic data and is shown to provide a better detection performance. The new detector is also tested on a real hyperspectral image.

International audience; Integrating spatial information into hyperspectral unmixing procedures has been shown to have a positive effect on the estimation of fractional abundances due to the inherent spatial-spectral duality in hyperspectral scenes. However, current research works that take spatial information into account are mainly focused on the linear mixing model. In this paper, we investigate how to incorporate spatial correlation into a nonlinear abundance estimation process. A nonlinear unmixing algorithm operating in reproducing kernel Hilbert spaces, coupled with a l1-type spatial regularization, is derived. Experiment results, with both synthetic and real hyperspectral images, illustrate the effectiveness of the proposed scheme.

International audience; Remote sensing data is not perfect and can be affected by some corruptions that may alter the data consistency, models created from the data and decisions made based on these models. Estimation error can diminish system efficiency in terms of classification accuracy. Accordingly, some learning approaches have adopted different ways to enhance their learning abilities from erroneous environments. Nevertheless, the existence of interpolation error can still introduce critical conflicting decisions. A more reasonable solution is introduced in this paper by incorporating this error into learning process. Instead of taking any unified theory of noise to evaluate the noise impacts separately, we try and to estimate this error and to incorporate it, cooperatively, into learning stage in order to enhance classification accuracy. Concretely, we present a new way to deal with reconstruction error when classifying 3D spectral signatures.

International audience; This paper studies a pixel by pixel nonlinearity detector for hyperspectral image analysis. The reflectances of linearly mixed pixels are assumed to be a linear combination of known pure spectral components (endmembers) contaminated by additive white Gaussian noise. Nonlinear mixing, however, is not restricted to any prescribed nonlinear mixing model. The mixing coefficients (abundances) satisfy the physically motivated sum-to-one and positivity constraints. The proposed detection strategy considers the distance between an observed pixel and the hyperplane spanned by the endmembers to decide whether that pixel satisfies the linear mixing model (null hypothesis) or results from a more general nonlinear mixture (alternative hypothesis). The distribution of this distance is derived under the two hypotheses. Closed-form expressions are then obtained for the probabilities of false alarm and detection as functions of the test threshold. The proposed detector is compared to another nonlinearity detector recently investigated in the literature through simulations using synthetic data. It is also applied to a real hyperspectral image.

International audience; Remote sensing data is not perfect and can be affected by some corruptions that may alter the data consistency, models created from the data and decisions made based on these models. Estimation error can diminish system efficiency in terms of classification accuracy. Accordingly, some learning approaches have adopted different ways to enhance their learning abilities from erroneous environments. Nevertheless, the existence of interpolation error can still introduce critical conflicting decisions. A more reasonable solution is introduced in this paper by incorporating this error into learning process. Instead of taking any unified theory of noise to evaluate the noise impacts separately, we try and to estimate this error and to incorporate it, cooperatively, into learning stage in order to enhance classification accuracy. Concretely, we present a new way to deal with reconstruction error when classifying 3D spectral signatures.

International audience; Nonlinear unmixing of hyperspectral images has generated considerable interest among researchers, as it may overcome some inherent limitations of the linear mixing model. In this paper, we formulate the problem of estimating abundances of a nonlinear mixture of hyperspectral data based on a new multi-kernel learning paradigm. Experiments are conducted using both synthetic and real images in order to illustrate the effectiveness of the proposed method.

In this paper we propose a novel and efficient model for compressed sensing of hyperspectral images. A large-size hyperspectral image can be subsampled by retaining only 3\% of its original size, yet robustly recovered using the new approach we present here. Our reconstruction approach is based on minimizing a convex functional which penalizes both the trace norm and the TV norm of the data matrix. Thus, the solution tends to have a simultaneous low-rank and piecewise smooth structure: the two important priors explaining the underlying correlation structure of such data. Through simulations we will show our approach significantly enhances the conventional compression rate-distortion tradeoffs. In particular, in the strong undersampling regimes our method outperforms the standard TV denoising image recovery scheme by more than 17dB in the reconstruction MSE.

International audience; The compression of hyperspectral images (HSIs) becomes recently very attractive issue for remote sensing applications because of their volumetric data. In this paper, an efficient method for joint hyperspectral image compression and noise reduction is presented. The proposed algorithm, based on Discrete Wavelet Transform and Tucker Decomposition (DWT- TD), exploits both the spectral and the spatial information in the images. The core idea behind our proposed technique is to apply TD on the DWT coefficients of spectral bands of HSIs. We use DWT to effectively separate HSIs into different sub images and TD to efficiently compact the energy of sub images. We evaluate the effect of the proposed method on real HSIs and also compare the results with the well-known compression methods. The obtained results show a better performance of the proposed method. Moreover, we show that the compression effects may be viewed to as an efficient noise reduction technique.

International audience; In hyperspectral images, pixels are mixtures of spectral components associated to pure materials. Although the linear mixture model is the most studied case, nonlinear models have been taken into consideration to overcome some limitations of the linear model. In this paper, nonlinear hyperspectral unmixing problem is studied through kernel-based learning theory. Endmember components at each band are mapped implicitly in a high feature space, in order to address the nonlinear interaction of photons. Experiment results with both synthetic and real images illustrate the effectiveness of the proposed scheme.

MCR-ALS is a resolution method that has been applied in many different fields, such as process analysis, environmental data and, recently, hyperspectral image analysis. In this context, the algorithm provides the distribution maps and the pure spectra of the image constituents from the sole information in the raw image measurement. Based on the distribution maps and spectra obtained, additional information can be easily derived, such as identification of constituents when libraries are available or quantitation within the image, expressed as constituent signal contribution. This work summarizes first the protocol followed for the resolution on two examples of kidney calculi, taken as representations of images with major and minor compounds, respectively. Image segmentation allows separating regions of images according to their pixel similarity and is also relevant in the biomedical field to differentiate healthy from non-healthy regions in tissues or to identify sample regions with distinct properties. Information on pixel similarity is enclosed not only in pixel spectra, but also in other smaller pixel representations, such as PCA scores. In this paper, we propose the use of MCR scores (concentration profiles) for segmentation purposes. K-means results obtained from different pixel representations of the data set are compared. The main advantages of the use of MCR scores are the interpretability of the class centroids and the compound-wise selection and preprocessing of the input information in the segmentation scheme.

This paper describes a new algorithm for hyperspectral image unmixing. Most of the unmixing algorithms proposed in the literature do not take into account the possible spatial correlations between the pixels. In this work, a Bayesian model is introduced to exploit these correlations. The image to be unmixed is assumed to be partitioned into regions (or classes) where the statistical properties of the abundance coefficients are homogeneous. A Markov random field is then proposed to model the spatial dependency of the pixels within any class. Conditionally upon a given class, each pixel is modeled by using the classical linear mixing model with additive white Gaussian noise. This strategy is investigated the well known linear mixing model. For this model, the posterior distributions of the unknown parameters and hyperparameters allow ones to infer the parameters of interest. These parameters include the abundances for each pixel, the means and variances of the abundances for each class, as well as a classification map indicating the classes of all pixels in the image. To overcome the complexity of the posterior distribution of interest, we consider Markov chain Monte Carlo methods that generate samples distributed according to the posterior of interest. The generated samples are then used for parameter and hyperparameter estimation. The accuracy of the proposed algorithms is illustrated on synthetic and real data., Comment: Manuscript accepted for publication in IEEE Trans. Geoscience and Remote Sensing

International audience; Stochastic watershed is a robust method to estimate the probability density function (pdf) of contours of a multi-variate image using MonteCarlo simulations of watersheds from random markers. The aim of this paper is to propose a stochastic watershed-based algorithm for segmenting hyperspectral images using a semi-supervised approach. Starting from a training dataset consisting in a selection of representative pixel vectors of each spectral class of the image, the algorithm calculate for each class a membership probability map (MPM). Then, the MPM of class k is considered as a regionalized density function which is used to simulate the random markers for the MonteCarlo estimation of the pdf of contours of the corresponding class k. This pdf favours the spatial regions of the image spectrally close to the class k. After applying the same technique to each class, a series of pdf are obtained for a single image. Finally, the pdf's can be segmented hierarchically either separately for each class or after combination, as a single pdf function. In the results, besides the generic spatial-spectral segmentation of hyperspectral images, the interest of the approach is also illustrated for target segmentation.

International audience; Support vector machines have been investigated with success for hyperspectral data classification. In this paper, we propose a new kernel to measure spectral similarity, called the angular kernel. We provide some of its properties, such as its invariance to illumination energy, as well as connection to previous work. Furthermore, we show that the performance of a classifier associated to the angular kernel is comparable to the Gaussian kernel, in the sense of universality. We derive a class of kernels based on the angular kernel, and study the performance on an urban classification task.

Hyperspectral remote sensing, also known as imaging spectroscopy, is a promising space technology regularly selected by agencies with regard to the exploration and observation of planets, to earth's geology or to the monitoring of the environment. It allows to collect for each pixel of a scene, the intensity of light energy reflected from planets as it varies across different wavelengths. More than one hundred spectels in the visible and near infra-red are typically recorded, making it possible to observe a continuous spectrum for each image cell. Usually, in space exploration, the analysis of these spectral signatures allows to retrieve the physical, chemical or mineralogical properties of surfaces and of atmospheres that may help to understand the geological and climatological history of planets. We propose in this paper a statistical method to evaluate the physical properties of surface materials on Mars from hyperspectral images collected by the OMEGA instrument aboard the Mars express spacecraft. The approach we develop is based on the estimation of the functional relationship F between some physical parameters and observed spectra. For this purpose, a database of synthetic spectra is generated by a physical radiative transfer model and used to estimate F. The high dimension of spectra is reduced by using Gaussian regularized sliced inverse regression (GRSIR) to overcome the curse of dimensionality and consequently the sensitivity of the inversion to noise (ill-conditioned problems). Compared with a naive spectrum matching approach such as the k-nearest neighbors algorithm, estimates are more accurate and realistic.