Annonce en lien avec l’Action/le Réseau : aucun

Laboratoire/Entreprise : ICUbe – Université de Strasbourg
Durée : 3 ans
Contact : gancarski@unistra.fr
Date limite de publication : 2020-04-31

Contexte :
Analyzing time series of remote sensing images using supervised methods requires that the classes sought be perfectly known and defined and that the expert be able to provide a learning data set that is sufficient in number and quality. Facing the difficulty of obtaining sufficient examples to efficient remote sensing time series analysis, we propose to develop an interactive method of collaborative clustering under constraints. The idea is to allow the expert to add “on the fly” constraints to guide the clustering process in order to produce clusters closer to the expert’s “intuitions” i.e. potential thematic classes. To do so, the expert will be helped by advice or proposals for new constraints issued by the method itself.

Sujet :
Nowadays, remote sensing images arrive massively and in almost continuous flow from the Sentinel constellation. This massive influx of temporal data should lead to major advances in various Earth and Environmental Science disciplines for the study and modelling of complex phenomena (agricultural or urban dynamics, deforestation, anthropogenic actions on biodiversity, etc.). Analyzing such time series using classical supervised methods requires that the thematic classes are perfectly known and defined. Unfortunately, in remote sensing domain, this assumption is not realistic. Indeed, the technological revolution of high-frequency image acquisition is still too recent for thematic knowledge to have adapted. Thus, there are currently no typologies (or nomenclatures) of changes that can really be used for this type of supervised analysis and therefore no associated quality learning data.
Faced with the difficulty of obtaining enough examples for the analysis of time series of remote sensing images, new clustering methods used constraints to guide the clustering process [1,3,4,5]. These unsupervised methods assume that this lack of knowledge can nevertheless be partially circumvented by using operable constraints (comparison, labelling or structural constraints). Such as constraints, which seem easier to define, can be used to guide the clustering process in order to produce clusters that are closer to the “intuitions” of the expert, i.e. potential thematic classes. In our team, we have developed SAMARAH an innovative method of collaborative interactive clustering under constraints [2] which allows the expert to add “on the fly” constraints.
Unfortunately, select new relevant constraints (object to be labelled, new constraint to apply…) that have positive impact on the current result, is often very difficult for the expert. Indeed, to define new constraints, the expert uses almost exclusively a visualization of the scene. Experiences have shown that, on the one hand, the expert focused on relatively large regions of the image and, on the other hand, he had no way of knowing whether the constraints he proposed were consistent with each other and relevant a priori. In fact, selecting new information is an important scientific lock, especially since it is essential to optimize the use of this new information from the expert. Indeed, if he does not see a rapid improvement of the solution following his help, he will quickly lose confidence in the system. But, paradoxically, the potential disruptions of the current solution should be limited in order not to disorient the expert. To this end, the expert must be assisted by advice or proposals for new constraints issued by the method in an active way [6,7].
The objective of this PhD is to study and implement mechanisms to propose potentially relevant constraints. This can be done, for example, by using differences in results due to the heterogeneity of methods in SAMARAH or by using a complexity measure, for example, based on trees of minimal weight to identify points at the boundaries between clusters and use them to define constraints.

Profil du candidat :

The positions are offered to both foreign and French students who hold a Master degree in computer science.

Formation et compétences requises :
The candidate must have good skills in data analysis and more particularly in supervised or unsupervised classification of time series. Skills in remote sensing image analysis is welcome.

Adresse d’emploi :
ICube
Pôle API – 300 bld Sébastien Brant
67400 Illkirch

Document attaché : Sujet_CNES_2020ENG.pdf