Les commentaires sont clos.

Optical and SAR remote sensing time series joint processing for snow, deforestation or sea ice monitoring.

20 Janvier 2022

Catégorie : Doctorant

The NASA MODIS program, composed of AQUA and TERRA satellites, provide numerous geo-referenced daily products such as snow cover, vegetation indices or sea ice extent. These products enable quick decision making for a variety of applications such as hydropower management or arctic navigation. The use of these product is nevertheless hindered by two major drawbacks :

  • The meteorological conditions, such as clouds, may impact the acquisition. In practice, the acquisition frequency is thus well above the days in some places.

  • The image resolution is about 500m, which limits the applications that need fine local information.

Other data sources are available freely such as the Copernicus satellites Sentinel-2 (S2) images (5 days repetition frequency, 10m resolution for the RGB bands), which acquisitions take place a few hours from the MODIS ones due to their heliosynchronous orbits. Thus, it is possible to get images with a resolution 50 times finer in a very short time interval. The Copernicus program also offers SAR images through the Sentinel-1 (S1) satellites constellation with two polarization channels, a 6-days repeat-time and a resolution of approximately 5x10 meters. The SAR sensors are not sensitive to clouds, but their geometry is very different to the optical images. SAR images have already demonstrated their capacity to measure biomass, detect wet snow, measure dry snow height and to classify the different types of sea ice.

On top of this large scale information, very local on-field measurement can be available such as snow gauge measures or sea-ice monitoring stations.


PhD goals:
This PhD aims to take advantage of all this information to predict regularly either the state of points on earth or the state change of points on earth, on a predefined georeferenced grid. Dependings on the application, the state can be the presence or absence of vegetation with the associated change described as deforestation or artificialisation, the presence or absence of snow and the associated change described as snowfall or snowmelt, or the type of sea ice. The PhD candidate will aim at proposing methodologies as generic as possible regarding the type of application.

Three research questions guide this work:

  • How to take into account the time information, while taking advantage of the spatial regularity of the observed processus. For example, snowmelt depends on the altitude of points, it will thus be interesting to take advantage of the information localized at lower or higher altitudes at a previous instant in time. Moreover, it will be important to keep the information of the order and the varying time delay between the different acquisitions.

  • How to make the best decision when some data is missing ? This can be the case when clouds impair optical acquisitions or when the acquisition dates are not matching. The developed methodologies must enable decision making by exploiting the integrality of available information.

  • What is the quantity of needed supervision ? For a large number of applications, establishing a ground truth is costly because it necessitates field work, or impossible because the events are too rare. Zero-shot learning approaches will thus be investigated for these situations.


These three research questions will help to choose the appropriate level in the methodology at which to merge the information. We are considering two main levels :

  • After a decision, assorted of an uncertainty measurement, has been taken by each modality. The regularization will be done a posteriori using methods such as Markov chains.

  • At the decision level, by using all modalities jointly to reduce the uncertainties. The regularity can then be learnt by algorithms such as deep neural networks.


The two levels of merging could be used jointly depending on the application.


Student profil:

The candidate holds an MSc diploma or equivalent in signal and image processing with strong knowledge in statistics. Skills in computer science (python,...) and an interest for physics are also required.


Period of employment:

The PhD will last 3 years from September/October 2022.


Supervising team:

Flora Weissgerber, Département traitement de l’information et systèmes, ONERA, Palaiseau

Jérôme Idier, LS2N, CNRS, Nantes

Sylvain Lobry, LIPADE, Université de Paris


How to candidate :

To candidate, send a CV and a cover letter/email at