Learning from Billions of Earth Pixels

Earth Observation has moved from an era of data scarcity to one of unexpected abundance. Where researchers once relied on a few carefully selected satellite scenes, open missions such as Landsat and Copernicus Sentinel, alongside innovative commercial missions, now deliver millions of images covering the entire planet at high spatial, spectral and temporal resolution. This growth has exposed the limits of traditional supervised learning approaches, which depend on costly and scarce labels. Self-supervised learning offers a new paradigm, learning directly from unlabelled imagery to construct rich, transferable representations of Earth's dynamics that go beyond simple label efficiency, enabling broad applicability across diverse domains.

This talk will guide you through the ongoing revolution in the field, showing several applications from mapping urban areas to rapid disaster assessment and forest monitoring, and reflecting on how the adoption of this technology across disciplines makes collaboration a foundational value of research in this field.