Seminars – 2023-2024

  • Date: THURSDAY, March 28th – 16h00
  • Room : VISIO
  • Speaker: Yann Cabanes (post-doct. Carleton University, Ottawa, Canada)
  • Title : Apprentissage dans les disques de Poincaré de séries temporelles complexes suivant un modèle autorégressif gaussien stationnaire centré : application au clustering de fouillis radar, à la détection et à la classification de drones
  • Abstract : L’objectif du travail présenté est l’étude de séries temporelles radar qui sont par nature des séries temporelles complexes centrées. Dans la première partie de cette présentation, nous souhaitons réaliser le clustering de fouillis radar, c’est-à-dire des données radar liées à l’environnement tels les mers, les forêts ou les champs environnants. Nous supposerons que les séries temporelles complexes observées suivent un modèle autorégressif gaussien stationnaire centré. De telles séries temporelles peuvent être représentées par leurs matrices de covariance qui sont des matrices Toeplitz hermitiennes définies positives. Elles peuvent également être représentées par les coefficients du modèle autorégressif. Certains coefficients autorégressifs appelés coefficients de réflexion sont de module strictement inférieur à 1 et déterminent entièrement le modèle autorégressif. Nous munirons cet espace de représentation d’une métrique riemannienne inspirée de la métrique de la géométrie de l’information sur les matrices hermitiennes définies positives. La variété riemannienne obtenue est une variété produit faisant intervenir plusieurs disques de Poincaré (autant que l’ordre du modèle autorégressif). Nous utiliserons alors l’algorithme des k-means dans cette variété riemannienne pour réaliser le clustering de fouillis radar. Dans la seconde partie de cette présentation, nous chercherons à faire de la détection et de la classification de drones à partir de séries temporelles radar. Les séries temporelles associées aux drones ne sont pas stationnaires, elles se distinguent par l’effet micro-Doppler induit par la rotation les hélices. Les séries temporelles complexes seront alors segmentées en fenêtres d’observation plus courtes. Pour chaque fenêtre, nous supposerons que la courte série temporelle observée est stationnaire et nous la représenterons par un point dans la variété riemannienne présentée dans la première partie. Nous obtiendrons alors une série temporelle riemannienne dont nous exploiterons les caractéristiques géométriques et statistiques pour faire de la détection et de la classification de drones.

———————————————

  • Date: THURSDAY, March 28th – 11h30
  • Room :  VISIO
  • Speaker: Marc Chaumont (Ass. professor HDR, LIRMM, Montpellier)
  • Title : Revisite d’une méthodologie « classique » pour l’analyse et la comparaison de détecteurs d’objets dans des séquences d’images
  • Abstract : Dans cet exposé, je présenterai tout d’abord mes activités de recherche en parcourant différents projets. Je présenterai ensuite une partie d’un projet européen dans lequel je suis engagé via la direction d’une thèse. Après avoir illustré ce projet, ainsi que les objectifs de la thèse, je vous parlerai d’une étape préliminaire, bien éloignée de l’objectif initial (!), et qui pourtant est non seulement importante mais également particulièrement intéressante d’un point de vue méthodologique/intellectuel. Ainsi, je vous parlerai de la construction d’un K-fold pour l’évaluation de détecteurs d’objets ! Bien que la technique de K-fold soit connue (dans le cadre de la classification) de tous « data-scientists » / « traiteur d’images » / statisticiens / etc, je ferai un petit rappel des diverses approches existantes, et je rappellerai dans quel cas l’utilisation d’un K-fold est pertinent. Bien évidemment, je vous exposerai l’approche que nous proposons afin de généraliser et d’adapter les techniques de K-fold de l’état-de-l’art à la tâche de détection d’objets.

———————————————

  • Date: WEDNESDAY, March 27th – 11h00
  • Room : VISIO
  • Speaker: Myriam BONTONOU (Post-doct., ENS, Lyon)
  • Title : Learning and Explaining Data Classification with Graph-Based Methods
  • Abstract : Graphs, concise representations of relationships between entities, are widely used in various fields such as vision, chemistry, or biology. This talk focuses on the use of graph neural networks for supervised classification of signals sharing a common graph structure. We will evaluate the performance of these networks in two application domains: molecular biology, where the graph integrates information on gene co-expression, and neuroimaging, where the graph represents neuronal connections between brain regions. Despite their potential, the performance of graph neural networks is not yet optimal. We will discuss current limitations and opportunities for improvement of these networks, as well as different ways in which graphs can contribute to a better understanding of neural networks.

———————————————

  • Date: MONDAY, March 25th – 11h00
  • Room : VISIO
  • Speaker: Patty COUPEAU (ATER Univ Angers, IUT GEII, LARIS)
  • Title : On the use of structural relationships based on deep learning and graphs for image analysis
  • Abstract : Deep learning for image analysis has grown considerably in recent decades, without exploiting all the information provided by images. My work focuses on the incorporation of high-level structural relationships, which correspond to observable relations between objects or regions of the image (e.g. spatial, photometric relationships), using graph neural networks (GNNs). Integrating this information allows for improved deep learning performance in image analysis and reduces the required amount of training data. The presentation will cover contributions to semantic image segmentation and image classification. Each contribution will be illustrated by a medical imaging application of central interest to my work, namely the study of early brain lesion in children, as well as a possible transposition to aerial imagery for environmental monitoring. Perspectives for these methods lie in the explainability of results, and the use of multi-view graphs, which will be useful, for example, for the efficient combination of spatial and temporal data from environmental observations.

———————————————

  • Date: THURSDAY, February 22d – 11h30
  • Room : A102, bât ENSIBS
  • Speaker: Stefano Polimena (Ph.D. student Department of Computer Science, University of Bari Aldo Moro, Italy)
  • Title :  Non-destructive contactless quality control in the agroalimentary supply chain
  • Abstract : The research of contactless and non-destructive quality assessment approaches for fruits and vegetables is of paramount importance in ensuring the integrity, safety, and reliability of products across the agroalimentary supply chain. It is crucial to know the quality of a product since it can provide benefits for both consumers, who are looking for healthy foods, and companies, to promote the design and distribution of products and services that minimize the environmental impact, preserve resources, and prioritize waste reduction. Recently, non-destructive computational methods have been proposed to replace the perception of human senses in order to estimate internal and external characteristics of the product. In this context, the application of machine learning and data mining techniques is useful to reduce labour costs, improve process efficiency and quality control reliability in the agroalimentary supply chain.  

———————————————

  • Date: THURSDAY, February 8th – 11h30
  • Room : A102, bât ENSIBS
  • Speaker: Iris Dumeur (Ph.D. student CESBIO, Toulouse)
  • Title : Paving the way toward foundation models for irregular and unaligned Satellite Image Time Series
  • Abstract : This presentation will depict our approach to filling existing gaps preventing the development of a foundation model for irregular and unaligned Satellite Image Time Series (SITS). First, we propose an ALIgned Sits Encoder (ALISE), which, while harnessing the spatial, spectral and temporal dimensions of the SITS, also produces fixed-dimensional embeddings from irregular and unaligned multi-year SITS. Second, inspired by recent advances in self-supervised learning, we suggest a multi-view self-supervised pre-training task combining a cross-reconstruction loss with latent space constraint losses. Eventually, results on downstream tasks show a single fully connected layer can successfully exploit aligned and frozen representations from ALISE. Specifically, on the PASTIS crop classification task, ALISE outperforms previous self-supervised methodologies: U-BARN [1] and Presto [2]. References
    [1] Iris Dumeur, Silvia Valero, and Jordi Inglada. Self-supervised spatio-temporal representation learning of satellite image time series. IEEE Journal of Selected Topics in Applied Earth Observations
    and Remote Sensing, nil(nil):1–18, 2024.
    [2] Gabriel Tseng, Ivan Zvonkov, Mirali Purohit, David Rolnick, and Hannah Rae Kerner. Lightweight, pre-trained transformers for remote sensing timeseries. ArXiv, abs/2304.14065, 2023.

———————————-

  • Date: WEDNESDAY, January 17th – 12h00
  • Room : A102, bât ENSIBS
  • Speaker: Maria LOBO, research scientist, and Jacques GAUTIER, assistant prof. (IGN, team GEOVIZ)
  • Title : Visualizing spatial data from 2D to 3D. Approaches and use cases.
  • Abstract:PART1 : Exploring transitions between map views Animation has been widely used in information visualization to depict changes between views, and to help users relate different representations. The design possibilities for these animations are very wide, but there are still few animations specifically designed for geographical representations. In this talk I will present two projects that explore different strategies to animate geographical representations. First, I will first present BAIA, a framework to create advanced animated transitions, called animation plans, between before-and-after satellite images. Second, I will present a study that compares different transitions between top-down 2D maps and 3D immersive views, to help users stay oriented when switching perspectives. PART2 : Visualization for analytical reasoning on spatial phenomenon. We present several projects involving Geovis team concerning the development of methods to improve visual analysis of meteorological and epidemiological data. First of all, the European project URCLIM, in which we proposed visual analysis approaches of simulated urban meteorological data and their relationships with urban morphology, through interactive 3D visualization allowing co-visualization of simulated 3D temperature data and urban morphological indicators. Then the ANR ORACLES project, in which we try to propose visualization approaches to allow a visual analysis of ensemble forecast scenarios of the coastal submersion phenomenon, by visualizing the disparities between multiple simulated scenarios of the submersion phenomenon and by co-visualizing submersion scenarios and associated meteorological-oceanic conditions. Finally, we present work from the COVISU project about visualization of spatio-temporal data related to the COVID-19 epidemic, describing the locations and dates of infection cases, for which we proposed two visualization approaches, one 2D representing the temporal distribution of cases inside spatial clusters of events and one 3D representing the evolution of the number of events according to a regular spatial grid, for the identification of spatio-temporal structures in the evolution of the epidemic (clusters, propagation axes).

————————————–

  • Date: WEDNESDAY, December  20th – 14:15
  • Room : salle de conférence de l’OSUR  (bâtiment 15 sur le campus Beaulieu), à Rennes + VISIO
  • Speaker: Mathilde Letard (Geoscience / LETG, Rennes) — Ph.D. defense
  • Title : Environmental knowledge extraction from topo-bathymetric lidar: machine learning and deep neural networks for point clouds and waveforms
  • Abstract : Land-water interfaces face escalating threats from climate change and human activities, necessitating systematic observation to comprehend and effectively address these challenges. Nevertheless, constraints associated with the presence of water hinder the uninterrupted observation of submerged and emerged areas. Topo-bathymetric lidar remote sensing emerges as a suitable solution, ensuring a continuous representation of land-water zones through 3D point clouds and 1D waveforms. However, fully harnessing the potential of this data requires tools specifically crafted to address its unique characteristics.  This thesis introduces methodologies for extracting environmental knowledge from topobathymetric lidar surveys. Initially, we introduce methods for classifying land and seabed covers using bi-spectral point clouds or waveform features. Subsequently, we employ deep neural networks for semantic segmentation, component detection and classification, and the estimation of water physical parameters based on bathymetric waveforms. Leveraging radiative transfer models, these approaches alleviate the need for manual waveform labeling, thereby enhancing waveform processing in challenging settings like extremely shallow or turbid waters.

————————————–

  • Date: THURSDAY, November, 16th – 10:00
  • Room : A103 (bat DSEG)
  • Speaker: Clément BONET (LMBA & IRISA-OBELIX, Univ. Bretagne-Sud, Vannes) — Ph.D. defense
  • Title : Leveraging Optimal Transport via Projections on Subspaces for Machine Learning Applications
  • Abstract : Optimal Transport has received many attention in Machine Learning as it allows comparing probability distributions by exploiting the geometry of the underlying space. However, in its original formulation, solving this problem suffers from a big computational burden. Thus, a big line of work consists at proposing alternatives to reduce this burden while still enjoying its properties. In this thesis, we focus on alternatives which use projections on subspaces. The main such alternative is the Sliced-Wasserstein distance, which we first propose to extend to Riemannian manifolds in order to use it in Machine Learning applications for which using such spaces has been shown beneficial in the recent years. We also study sliced distances between positive measures in the so-called unbalanced OT problem. Back to the original Euclidean Sliced-Wasserstein distance between probability measures, we study the dynamic of gradient flows when endowing the space with this distance in place of the usual Wasserstein distance. Then, we investigate the use of the Busemann function, a generalization of the inner product in metric space, in the space of probability measures. Finally, we extend the subspace detour approach to incomparable spaces using the Gromov-Wasserstein distance.

————————————–

  • Date: FRIDAY, Septembre, 22d – 11:30
  • Room : A102 (bat ENSIBS)
  • Speaker: Aimi OKABAYASHI (EN SAT, Paris)
  • Title : Diffusion models for the super-resolution of satellite image time series
  • Abstract : Satellite imagery is a major tool for geoscience. However, there is a dilemma on whether to deploy systems that acquire many images at high frequency but with a low spatial resolution or systems with high spatial resolution but few revisits. For example, on the one hand, Sentinel-2 offers a high temporal frequency but a spatial resolution limited to 10m/px, which is not enough for some applications such as urban mapping of buildings, roads, or sparse vegetation. On the other hand, SPOT-6 delivers a high spatial resolution of 1.5m/px but is only used for commercial purposes, and distributes only once a year a high-resolution cloud-free mosaic of the French metropolitan area. To benefit from the high temporal frequency of Sentinel-2 at a high spatial resolution, we propose to apply a super-resolution technique. Deep learning approaches have been frequently explored over the years, starting from CNNs to GANs. Recent work in Super-Resolution demonstrated impressive results with diffusion models on synthetic data, i.e., the high- and low-spatial resolution image pairs (HR, LR) are artificially created by downsampling the target HR image to obtain the LR image. In this talk, I will present the performance of diffusion models on satellite data, coming from two different sources (Sentinel-2 and SPOT-6). Moreover, I will explore how diffusion models could benefit from the temporal information in the time series to construct a better super-resolved image.

Comments are closed.