Les commentaires sont clos.

Stage M2@ IGN: Spatio-Temporal Domain Adaptation for Country-Scale Land-Cover Prediction

18 Novembre 2021

Catégorie : Stagiaire

We propose an internship as part of the TerrIA project to develop machine learning tools for large-scale land cover prediction. A particularily interesting challenges in terms of applciation and method is to develop methods that are robust to the multiple domain shifts affecting country-scale geographical data: different bioclimatic conditions, land use, time and weather during the acquisition, etc..

This internship consists in an investigation of the relevance of different state-of-the-art domain adaptation method to the case of large-scale land cover, and the release of an open-source dataset and associated publciation.


Spatio-Temporal Domain Adaptation for Country-Scale Land-Cover Prediction

Internship proposal, 6 months

  • Laboratory: équipe STRUDEL, Laboratoire LaSTIG (IGN/Univ. Gustave Eiffel)
  • Localisation: IGN: Saint Mandé, France
  • Supervision: Loic Landrieu, PhD; Sebastien Giordano, PhD, Nicolas David
  • Remuneration: 513 euros / mois
  • Starting Date: April 2022, up to 6 month duration
  • InECCV, 2KeyWords: Deep Learning, Domain Shift, Large-Scale, High-Resolution, Open-Source, Environment Monitoring


IGN is the public institution in charge of the production and distribution of geographical information in France. As part of a push towards full open-access (the “géo-communs” policy), the Terr-IA project aims to produce pixel-precise annotate over 1400 km2 of high-definition aerial images (20cm/pixel) accross 50 départements. This is the first step towards fully automated digitization of a national and regurlaly updated land cover/use database from multi-source remote sensing, with the aim of monitoring anthropegenic environmental impact (deforestation, impervious surfaces, urban sprawl, etc.). All data and code will be open-source, and aim to reach the international research community.
This dataset presents a major opportunity in terms of machine learning research. Indeed, differences in distribution between training and test sets are known to severely impact the performance of classification algorithms [4]. Such distribution-shift are typically met when considering geographical data acquired from regions with different bioclimatic conditions or land usage (urban/rural/natural). The date and even time of acquisition can have a profound impact on the captured data as well and must be accounted for.


The goal of this internship is to evaluate the relevance of state-of-the-art approaches for training a neural network from data from diverse acquisition domains (spatial and temporal), and able to generalize to unseen areas and conditions. We will curate a sub-dataset from the Terr-IA annotated data to evaluate different domain adaptation approaches. We will investigate the performance of metalearning [5], adversarial approaches [7], and distribution regularizers [6, 1] for spatio-temporal domain adaptation. If successful, the intern will write an article to present the dataset and our analysis. Observations about what works and what does not may lead us to modify and improve SOTA approaches
to fit the problem at hand.

The tentative planning of the internship is as follows:

  • Month 1-2. Familiarization with the annotated dataset, curation of an illustrative sub-dataset. Choice and implementation of baselines for land-use prediction.
  • Month 3-4. Bibliography and implementation of relevant domain adaptation methods.
  • Month 5-6. Running numerical experiments, writing the article.
  • Bonus. Improving state-of-the-art methods according to our analysis of the performance of existing approaches.


The intern will have a privileged opportunity to postulate to a PhD extending this project, and whose scope includes(i) impact of multispectral and radar satellite time series [3, 2], (ii) integration with the LiDAR-HD acquisition project (country-scale 3D acquisition), (iii) prediction of vectorized data, (iv) release of country-scale open-access datasets.


  • Master 2 student in computer science, applied mathematics, or remote sensing.
  • Familiarity with computer vision, machine learning, and deep learning.
  • Mastery of Python, familiarity with PyTorch;
  • Curiosity, rigor, motivation;
  • (Optional) Familiarity with domain adaptation methods;
  • (Optional) Experienced with aerial/satellite sensor technology and land-use prediction models.


Send a CV and a short letter of purpose (∼20 ligns max) stating your interest for this internship and the relevance of your experience to and

[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019.
[2] Vivien Sainte Fare Garnot and Loic Landrieu. Panoptic segmentation of satellite image time series with convolutional temporal attention networks. In ICCV, 2021.
[3] Vivien Sainte Fare Garnot, Loic Landrieu, Sebastien Giordano, and Nesrine Chehata. Satellite image time series classification with pixel-set encoders and temporal self-attention. In CVPR, 2020.
[4] Pang Wei Koh, Shiori Sagawa, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, et al. Wilds: A benchmark of in-the-wild distribution shifts. In ICLR, 2021.
[5] Marc Rußwurm, Sherrie Wang, Marco Korner, and David Lobell. Meta-learning for few-shot land cover classification. In CVPR Workshop, 2020.
[6] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, 2016.
[7] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017.