Domain Adaptation in remote sensing: increasing the portability of land-cover classifiers

Among the types of remote sensing acquisitions, optical images are certainly one of the most widely relied upon data sources for Earth observation. They provide detailed measurements of the electromagnetic radiation reflected or emitted by each pixel in the scene. Through a process termed supervised land-cover classification, this allows to automatically yet accurately distinguish objects at the surface of our planet. In this respect, when producing a land-cover map of the surveyed area, the availability of training examples representative of each thematic class is crucial for the success of the classification procedure. However, in real applications, due to several constraints on the sample collection process, labeled pixels are usually scarce. When analyzing an image for which those key samples are unavailable, a viable solution consists in resorting to the ground truth data of other previously acquired images. This option is attractive but several factors such as atmospheric, ground and acquisition conditions can cause radiometric differences between the images, hindering therefore the transfer of knowledge from one image to another. The goal of this Thesis is to supply remote sensing image analysts with suitable processing techniques to ensure a robust portability of the classification models across different images. The ultimate purpose is to map the land-cover classes over large spatial and temporal extents with minimal ground information. To overcome, or simply quantify, the observed shifts in the statistical distribution of the spectra of the materials, we study four approaches issued from the field of machine learning. First, we propose a strategy to intelligently sample the image of interest to collect the labels only in correspondence of the most useful pixels. This iterative routine is based on a constant evaluation of the pertinence to the new image of the initial training data actually belonging to a different image. Second, an approach to reduce the radiometric differences among the images by projecting the respective pixels in a common new data space is presented. We analyze a kernel-based feature extraction framework suited for such problems, showing that, after this relative normalization, the cross-image generalization abilities of a classifier are highly increased. Third, we test a new data-driven measure of distance between probability distributions to assess the distortions caused by differences in the acquisition geometry affecting series of multi-angle images. Also, we gauge the portability of classification models through the sequences. In both exercises, the efficacy of classic physically- and statistically-based normalization methods is discussed. Finally, we explore a new family of approaches based on sparse representations of the samples to reciprocally convert the data space of two images. The projection function bridging the images allows a synthesis of new pixels with more similar characteristics ultimately facilitating the land-cover mapping across images.

Ensemble classifiers in remote sensing: a comparative analysis

Dissertation submitted in partial fulfillment of the requirements for the Degree of Master of Science in Geospatial Technologies.

Kernel-based methods for change detection in remote sensing images

Nowadays, the joint exploitation of images acquired daily by remote sensing instruments and of images available from archives allows a detailed monitoring of the transitions occurring at the surface of the Earth. These modifications of the land cover generate spectral discrepancies that can be detected via the analysis of remote sensing images. Independently from the origin of the images and of type of surface change, a correct processing of such data implies the adoption of flexible, robust and possibly nonlinear method, to correctly account for the complex statistical relationships characterizing the pixels of the images. This Thesis deals with the development and the application of advanced statistical methods for multi-temporal optical remote sensing image processing tasks. Three different families of machine learning models have been explored and fundamental solutions for change detection problems are provided. In the first part, change detection with user supervision has been considered. In a first application, a nonlinear classifier has been applied with the intent of precisely delineating flooded regions from a pair of images. In a second case study, the spatial context of each pixel has been injected into another nonlinear classifier to obtain a precise mapping of new urban structures. In both cases, the user provides the classifier with examples of what he believes has changed or not. In the second part, a completely automatic and unsupervised method for precise binary detection of changes has been proposed. The technique allows a very accurate mapping without any user intervention, resulting particularly useful when readiness and reaction times of the system are a crucial constraint. In the third, the problem of statistical distributions shifting between acquisitions is studied. Two approaches to transform the couple of bi-temporal images and reduce their differences unrelated to changes in land cover are studied. The methods align the distributions of the images, so that the pixel-wise comparison could be carried out with higher accuracy. Furthermore, the second method can deal with images from different sensors, no matter the dimensionality of the data nor the spectral information content. This opens the doors to possible solutions for a crucial problem in the field: detecting changes when the images have been acquired by two different sensors.

Recent advances in remote sensing image processing

Remote sensing image processing is nowadays a mature research area. The techniques developed in the field allow many real-life applications with great societal value. For instance, urban monitoring, fire detection or flood prediction can have a great impact on economical and environmental issues. To attain such objectives, the remote sensing community has turned into a multidisciplinary field of science that embraces physics, signal theory, computer science, electronics, and communications. From a machine learning and signal/image processing point of view, all the applications are tackled under specific formalisms, such as classification and clustering, regression and function approximation, image coding, restoration and enhancement, source unmixing, data fusion or feature selection and extraction. This paper serves as a survey of methods and applications, and reviews the last methodological advances in remote sensing image processing.

Rank Aggregation for Pattern Classifier Selection in Remote Sensing Images

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES)

TECHNOLOGICAL ADVANCEMENT IN REMOTE SENSING AND INNOVATIVE APPLICATION TO VOLCANOLOGY

The integration of remote monitoring techniques at different scales is of crucial importance for monitoring of volcanoes and assessment of the associated hazard. In this optic, technological advancement and collaboration between research groups also play a key role. Vhub is a community cyberinfrastructure platform designed for collaboration in volcanology research. Within the Vhub framework, this dissertation focuses on two research themes, both representing novel applications of remotely sensed data in volcanology: advancement in the acquisition of topographic data via active techniques and application of passive multi-spectral satellite data to monitoring of vegetated volcanoes. Measuring surface deformation is a critical issue in analogue modelling of Earth science phenomena. I present a novel application of the Microsoft Kinect sensor to measurement of vertical and horizontal displacements in analogue models. Specifically, I quantified vertical displacement in a scaled analogue model of Nisyros volcano, Greece, simulating magmatic deflation and inflation and related surface deformation, and included the horizontal component to reconstruct 3D models of pit crater formation. The detection of active faults around volcanoes is of importance for seismic and volcanic hazard assessment, but not a simple task to be achieved using analogue models. I present new evidence of neotectonic deformation along a north-south trending fault from the Mt Shasta debris avalanche deposit (DAD), northern California. The fault was identified on an airborne LiDAR campaign of part of the region interested by the DAD and then confirmed in the field. High resolution LiDAR can be utilized also for geomorphological assessment of DADs, and I describe a size-distance analysis to document geomorphological aspects of hummock in the Shasta DAD. Relating the remote observations of volcanic passive degassing to conditions and impacts on the ground provides an increased understanding of volcanic degassing and how satellite-based monitoring can be used to inform hazard management strategies in nearreal time. Combining a variety of satellite-based spectral time series I aim to perform the first space-based assessment of the impacts of sulfur dioxide emissions from Turrialba volcano, Costa Rica, on vegetation in the surrounding environment, and establish whether vegetation indices could be used more broadly to detect volcanic unrest.

Distributed Visualization Systems in Remote Sensing Data Processing Grid

Implementation of GEOSS/GMES initiative requires creation and integration of service providers, most of which provide geospatial data output from Grid system to interactive user. In this paper approaches of DOS- centers (service providers) integration used in Ukrainian segment of GEOSS/GMES will be considered and template solutions for geospatial data visualization subsystems will be suggested. Developed patterns are implemented in DOS center of Space Research Institute of National Academy of Science of Ukraine and National Space Agency of Ukraine (NASU-NSAU).

REFLECT : logiciel de restitution des réflectances au sol pour l’amélioration de la qualité de l'information extraite des images satellitales à haute résolution spatiale

RÉSUMÉ - Les images satellitales multispectrales, notamment celles à haute résolution spatiale (plus fine que 30 m au sol), représentent une source d’information inestimable pour la prise de décision dans divers domaines liés à la gestion des ressources naturelles, à la préservation de l’environnement ou à l’aménagement et la gestion des centres urbains. Les échelles d’étude peuvent aller du local (résolutions plus fines que 5 m) à des échelles régionales (résolutions plus grossières que 5 m). Ces images caractérisent la variation de la réflectance des objets dans le spectre qui est l’information clé pour un grand nombre d’applications de ces données. Or, les mesures des capteurs satellitaux sont aussi affectées par des facteurs « parasites » liés aux conditions d’éclairement et d’observation, à l’atmosphère, à la topographie et aux propriétés des capteurs. Deux questions nous ont préoccupé dans cette recherche. Quelle est la meilleure approche pour restituer les réflectances au sol à partir des valeurs numériques enregistrées par les capteurs tenant compte des ces facteurs parasites ? Cette restitution est-elle la condition sine qua non pour extraire une information fiable des images en fonction des problématiques propres aux différents domaines d’application des images (cartographie du territoire, monitoring de l’environnement, suivi des changements du paysage, inventaires des ressources, etc.) ? Les recherches effectuées les 30 dernières années ont abouti à une série de techniques de correction des données des effets des facteurs parasites dont certaines permettent de restituer les réflectances au sol. Plusieurs questions sont cependant encore en suspens et d’autres nécessitent des approfondissements afin, d’une part d’améliorer la précision des résultats et d’autre part, de rendre ces techniques plus versatiles en les adaptant à un plus large éventail de conditions d’acquisition des données. Nous pouvons en mentionner quelques unes : - Comment prendre en compte des caractéristiques atmosphériques (notamment des particules d’aérosol) adaptées à des conditions locales et régionales et ne pas se fier à des modèles par défaut qui indiquent des tendances spatiotemporelles à long terme mais s’ajustent mal à des observations instantanées et restreintes spatialement ? - Comment tenir compte des effets de « contamination » du signal provenant de l’objet visé par le capteur par les signaux provenant des objets environnant (effet d’adjacence) ? ce phénomène devient très important pour des images de résolution plus fine que 5 m; - Quels sont les effets des angles de visée des capteurs hors nadir qui sont de plus en plus présents puisqu’ils offrent une meilleure résolution temporelle et la possibilité d’obtenir des couples d’images stéréoscopiques ? - Comment augmenter l’efficacité des techniques de traitement et d’analyse automatique des images multispectrales à des terrains accidentés et montagneux tenant compte des effets multiples du relief topographique sur le signal capté à distance ? D’autre part, malgré les nombreuses démonstrations par des chercheurs que l’information extraite des images satellitales peut être altérée à cause des tous ces facteurs parasites, force est de constater aujourd’hui que les corrections radiométriques demeurent peu utilisées sur une base routinière tel qu’est le cas pour les corrections géométriques. Pour ces dernières, les logiciels commerciaux de télédétection possèdent des algorithmes versatiles, puissants et à la portée des utilisateurs. Les algorithmes des corrections radiométriques, lorsqu’ils sont proposés, demeurent des boîtes noires peu flexibles nécessitant la plupart de temps des utilisateurs experts en la matière. Les objectifs que nous nous sommes fixés dans cette recherche sont les suivants : 1) Développer un logiciel de restitution des réflectances au sol tenant compte des questions posées ci-haut. Ce logiciel devait être suffisamment modulaire pour pouvoir le bonifier, l’améliorer et l’adapter à diverses problématiques d’application d’images satellitales; et 2) Appliquer ce logiciel dans différents contextes (urbain, agricole, forestier) et analyser les résultats obtenus afin d’évaluer le gain en précision de l’information extraite par des images satellitales transformées en images des réflectances au sol et par conséquent la nécessité d’opérer ainsi peu importe la problématique de l’application. Ainsi, à travers cette recherche, nous avons réalisé un outil de restitution de la réflectance au sol (la nouvelle version du logiciel REFLECT). Ce logiciel est basé sur la formulation (et les routines) du code 6S (Seconde Simulation du Signal Satellitaire dans le Spectre Solaire) et sur la méthode des cibles obscures pour l’estimation de l’épaisseur optique des aérosols (aerosol optical depth, AOD), qui est le facteur le plus difficile à corriger. Des améliorations substantielles ont été apportées aux modèles existants. Ces améliorations concernent essentiellement les propriétés des aérosols (intégration d’un modèle plus récent, amélioration de la recherche des cibles obscures pour l’estimation de l’AOD), la prise en compte de l’effet d’adjacence à l’aide d’un modèle de réflexion spéculaire, la prise en compte de la majorité des capteurs multispectraux à haute résolution (Landsat TM et ETM+, tous les HR de SPOT 1 à 5, EO-1 ALI et ASTER) et à très haute résolution (QuickBird et Ikonos) utilisés actuellement et la correction des effets topographiques l’aide d’un modèle qui sépare les composantes directe et diffuse du rayonnement solaire et qui s’adapte également à la canopée forestière. Les travaux de validation ont montré que la restitution de la réflectance au sol par REFLECT se fait avec une précision de l’ordre de ±0.01 unités de réflectance (pour les bandes spectrales du visible, PIR et MIR), même dans le cas d’une surface à topographie variable. Ce logiciel a permis de montrer, à travers des simulations de réflectances apparentes à quel point les facteurs parasites influant les valeurs numériques des images pouvaient modifier le signal utile qui est la réflectance au sol (erreurs de 10 à plus de 50%). REFLECT a également été utilisé pour voir l’importance de l’utilisation des réflectances au sol plutôt que les valeurs numériques brutes pour diverses applications courantes de la télédétection dans les domaines des classifications, du suivi des changements, de l’agriculture et de la foresterie. Dans la majorité des applications (suivi des changements par images multi-dates, utilisation d’indices de végétation, estimation de paramètres biophysiques, …), la correction des images est une opération cruciale pour obtenir des résultats fiables. D’un point de vue informatique, le logiciel REFLECT se présente comme une série de menus simples d’utilisation correspondant aux différentes étapes de saisie des intrants de la scène, calcul des transmittances gazeuses, estimation de l’AOD par la méthode des cibles obscures et enfin, l’application des corrections radiométriques à l’image, notamment par l’option rapide qui permet de traiter une image de 5000 par 5000 pixels en 15 minutes environ. Cette recherche ouvre une série de pistes pour d’autres améliorations des modèles et méthodes liés au domaine des corrections radiométriques, notamment en ce qui concerne l’intégration de la FDRB (fonction de distribution de la réflectance bidirectionnelle) dans la formulation, la prise en compte des nuages translucides à l’aide de la modélisation de la diffusion non sélective et l’automatisation de la méthode des pentes équivalentes proposée pour les corrections topographiques.

Datasets related to in-land water for limnology and remote sensing applications: distance-to-land, distance-to-water, water-body identifier and lake-centre co-ordinates

Datasets containing information to locate and identify water bodies have been generated from data locating static-water-bodies with resolution of about 300 m (1/360 deg) recently released by the Land Cover Climate Change Initiative (LC CCI) of the European Space Agency. The LC CCI water-bodies dataset has been obtained from multi-temporal metrics based on time series of the backscattered intensity recorded by ASAR on Envisat between 2005 and 2010. The new derived datasets provide coherently: distance to land, distance to water, water-body identifiers and lake-centre locations. The water-body identifier dataset locates the water bodies assigning the identifiers of the Global Lakes and Wetlands Database (GLWD), and lake centres are defined for in-land waters for which GLWD IDs were determined. The new datasets therefore link recent lake/reservoir/wetlands extent to the GLWD, together with a set of coordinates which locates unambiguously the water bodies in the database. Information on distance-to-land for each water cell and the distance-to-water for each land cell has many potential applications in remote sensing, where the applicability of geophysical retrieval algorithms may be affected by the presence of water or land within a satellite field of view (image pixel). During the generation and validation of the datasets some limitations of the GLWD database and of the LC CCI water-bodies mask have been found. Some examples of the inaccuracies/limitations are presented and discussed. Temporal change in water-body extent is common. Future versions of the LC CCI dataset are planned to represent temporal variation, and this will permit these derived datasets to be updated.

Distributed processing of large remote sensing images using MapReduce - A case of Edge Detection

Dissertation submitted in partial fulfillment of the requirements for the Degree of Master of Science in Geospatial Technologies.

Structured output SVM for remote sensing image classification

In the recent years, kernel methods have revealed very powerful tools in many application domains in general and in remote sensing image classification in particular. The special characteristics of remote sensing images (high dimension, few labeled samples and different noise sources) are efficiently dealt with kernel machines. In this paper, we propose the use of structured output learning to improve remote sensing image classification based on kernels. Structured output learning is concerned with the design of machine learning algorithms that not only implement input-output mapping, but also take into account the relations between output labels, thus generalizing unstructured kernel methods. We analyze the framework and introduce it to the remote sensing community. Output similarity is here encoded into SVM classifiers by modifying the model loss function and the kernel function either independently or jointly. Experiments on a very high resolution (VHR) image classification problem shows promising results and opens a wide field of research with structured output kernel methods.

Active learning methods for remote sensing image classification

In this paper, we propose two active learning algorithms for semiautomatic definition of training samples in remote sensing image classification. Based on predefined heuristics, the classifier ranks the unlabeled pixels and automatically chooses those that are considered the most valuable for its improvement. Once the pixels have been selected, the analyst labels them manually and the process is iterated. Starting with a small and nonoptimal training set, the model itself builds the optimal set of samples which minimizes the classification error. We have applied the proposed algorithms to a variety of remote sensing data, including very high resolution and hyperspectral images, using support vector machines. Experimental results confirm the consistency of the methods. The required number of training samples can be reduced to 10% using the methods proposed, reaching the same level of accuracy as larger data sets. A comparison with a state-of-the-art active learning method, margin sampling, is provided, highlighting advantages of the methods proposed. The effect of spatial resolution and separability of the classes on the quality of the selection of pixels is also discussed.