Tristan Postadjian

Au MATIS depuis 2015

Téléphone actuel : 33 1 43 98 80 00 + 7102

Contact : tristan.postadjian(at)

Axe de recherche RISOTO

Sujet de recherche


Towards a full France land cover with very high resolution satellite images.

Land cover knowledge is necessary for a territory description, and to help local authorities and government to make decisions related to it (protection of farmlands, limitation of space consumption, ecology issues, ...). To achieve that, land cover database is being made with a classification that will fit the needs of the end users. With this in mind, the French Mapping Agency builds a large scale land cover (OCS-GE), using existing databases (aggregation) and spatial analysis tools.

Images acquired recently by launched satellites dedicated to the Earth observation, Spot 6/7, allow the production of an annual very high resolution orthoimage map.

The new classification is performed on these images and is based on less resolved classifications, in particular the one computed by the CESBIO : the goal is to preserve the classes it detected thanks to the multitemporal analysis (comparison between several images acquired at different dates) and sharpen it with new classes. Hence, a hierarchical approach is used. Deep-Learning methods (Convolutional Neural Networks) are explored as serious alternative to more "classic" algorithm (SVM, Random Forests) for this part, regarding to the images at our disposal and the big data dimension of the topic.

Three main challenges :
  • The classification uses databases that have variable qualities ; therefore, it has to be taken into account before processing, and it would be ideally possible to qualify those databases prior to any computation.
  • An other important scope is the scale of study : the whole metropolitan France. Parallel computing has to be performed, and then, the France must be partitioned optimally.
  • Finally, classes used during the process are different from those that actually interest the end users. Thus, relevant bridges must be done between them.

Personal webpage


Thèse en cours en depuis 2015 (Université Paris Est - Ecole Doctorale MSTIC):
Vers une occupation du sol France entière à très haute résolution
Dirigée par Clément Mallet, encadrement : Arnaud Le_Bris et Hichem Sahbi.

De 2012 à 2015 : Cycle Ingénieur à l’École Nationale des Sciences Géographiques


Articles de conférences avec comité de lecture

T. Postadjian, C. Mallet, A. Le_Bris, H. Sahbi. Investigating the potential of deep neural networks for large-scale classification of very high resolution satellite images. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2017.

Site internet de la recherche à l'IGN