Les commentaires sont clos.

[StageM2] Representation Learning for Geographic Spatio-Temporal Generalisation - ICube, Strasbourg and CNES

16 Novembre 2022

Catégorie : Stagiaire

Time-series are becoming prevalent in many fields, particularly when monitoring environmental changes of the Earth's surface in the long term (climate change, urbanisation, etc), medium term (annual crop cycle, etc) or short term (earthquakes, floods, etc). With the current and future satellite constellations satellite image time-series (SITS) expand remote sensing's impact.


Detailed subject:

The project's goal is to develop domain invariant representations using deep learning for SITS analysis. Such methods will enable geographic generalisation, which consists of reusing information from the analysis of one geographic area to analyse others by using, or not, the same sensors, as proposed in [5]. Current approaches work for single images because they generally originate from the computer vision community. The internship will start the evaluation of the state-of-the-art and to implement and extend approaches already developed in ICube [5,6]. Current work on domain adaptation (DA) for time-series uses either weak supervision [1] or attention-based mechanisms [2,3] for classification or focus on the related problem of time-series forecasting [4]. However, none of these approaches tackle the problem of learning DIRs that can be applied to several geographical locations simultaneously. The work has two benefits: on the one hand, to reduce the burden of ground truth collection when sensors of different characteristics are used; and on the other to exploit the information contained in each data modality to learn representations that are more robust and general, i.e. to detect crops, land cover evolution, etc in different countries that exhibit different characteristics. Your contributions will be part of the global work of the SDC researchers and will be validated through the partnership with CNES and potential collaboration with Tour du Valat. SDC's aim is to propose and implement new generic methods and tools to exploit large sets of reference data from one domain/modality (sufficient to train an accurate detector) to train a multi-modal/domain detector that can be applied to imagery taken from another sensor for which there exists no reference data. As such, the work tackles key problems in many machine learning & computer vision applications.

To apply, send an email to Thomas Lampert (, be sure to include your C.V., transcript of your grades, your rank in class if possible, and a cover letter to let us know why you think you would be a good fit.