ISTerre seminar

---

Self-supervised representation learning for remote sensing: an introduction and recent contribution

Friday 17 January 2025 - 11h00

Jules Bourcier - ISTerre
--- 
Remote sensing is a crucial tool for monitoring and understanding our planet and human activities at a global scale. Deep neural networks, which can learn powerful semantic representations from data, are one way to make significant progress in remote sensing image analysis for Earth observation tasks, such as land use and land cover mapping. However, the limited availability of labeled data for these tasks poses a significant challenge, as traditional supervised learning methods require big labeled datasets to (pre)train models. In this talk, I will introduce self-supervised learning of visual representations, which intends to learn deep feature extractors from unlabeled images that can then be transferred to tackle downstream tasks with limited annotated data. Such methods are of great interest for remote sensing due to the abundance of data and scarcity of labels. The specific charateristics and challenges of remote sensing data form great opportunities for improving the efficiency and accuracy of self-supervised methods. I will present one recent contribution, where we introduce a new multimodal pretraining framework leveraging the metadata of satellite images as a direct supervision signal.

---

Organizing team : Cycle sismique et déformations transitoires

Salle Dolomieu, Maison des Géosciences, 38400 Saint Martin d'Hères