Analyse de Séries Temporelles Massives d'Images Satellitaires : Applications à la Cartographie des Écosystèmes
Spécialité : Mathématiques Appliquées
13/12/2021 - 14:00 Alexandre Constantin (Université Grenoble Alpes) Amphithéâtre, Bâtiment IMAG-UGA, 700 Avenue Centrale, 38400 Saint-Martin-d'Hères
Mots clé :
- Time Series
- Optimization
- Big Data
- Earth Observation
This thesis takes place in the context of the processing of the data from Sentinel-2 mission. This mission, initiated by the European Space Agency and launched in 2017, produces an unprecedented amount of Satellite Image Time-Series (SITS). Among the key analyses of these images, this thesis focuses on the classification task, i.e. land use or land cover maps that can be produced using spectro-temporal aspect of the Sentinel-2 SITS. Two main difficulties are identified in this thesis for the process of Sentinel-2 SITS. First, the unprecedented amount of data requires both scalable classifiers and code optimization techniques (such as parallel processing). Second, the acquisition noise (clouds, shadows) combined with the temporal aspect results in irregular and unevenly sampled time-series. Conventional approaches re-sample time-series to a set of time stamps, then they use machine learning techniques to classify vectors at a large-scale (national scale). The main disadvantage of this two-step processing approach is that it increases the number of operations applied to the SITS, implying a more difficult transition to massive amount of data. To a lower extent, the re-sampling step may slightly alter the temporal characteristics of the data. This thesis contributions are the following. We introduce a novel model-based approach with the ability to classify irregularly sampled time-series based on a mixture of multivariate Gaussian processes. A two-step approach has been used, by defining on one hand a model of uni-variate time-series, independent from the spectral wavelength point of view, then by considering on the second hand both spectral and temporal information from SITS. These models allow jointly a reconstruction of unobserved or noisy data. Estimation of both models has been implemented using a parallelized python code to be scalable to large-scale data-sets. The two models are evaluated numerically on Sentinel-2 SITS in terms of classification and reconstruction accuracy and are compared with conventional approaches. Analyses of the results illustrate the relevance of the two models and the benefit of using interpretable parametric models.
Directeurs:
- Stéphane Girard (Inria Grenoble Rhône-Alpes )
Raporteurs:
- Gabriele Moser (Università Degli Studi di Genova )
- Julien Jacques (Université Lyon 2 - A & L Lumière )
Examinateurs:
- Marie Chabert (INP - ENSEEIHT Toulouse )
- Lionel Bombrun (Bordeaux Sciences Agro )