Scope

Earth observation is a research subfield of remote sensing that uses satellite imagery for various applications in agriculture, environmental monitoring, weather forecasting, geology, risk management, security and public health using satellite observations. The field is ideally suited for various deep learning methods, which primarily rely on well-organized datasets of labeled samples. However, labels in the field of Earth observation are difficult to obtain (i.e., limited) and may come from various indirect sources that are not fully consistent with the collected satellite imagery (i.e., noisy or uncertain). Furthermore, the distribution of labeled patterns is often quite unbalanced, with only a few training patterns available for a given label. The main objective of the proposed project is to investigate the combined effects of self-supervised and multiple robust deep learning techniques in the context of Earth observation applications.

Overview of the project idea

Workpackages

• WP1: Self-Supervised Learning for Efficient Representations
• WP2: Robust Learning Based on Representations
• WP3: Robust Learning in Earth Observation Tasks
• WP4: Management & Dissemination