We show that powerful image-based AI models, originally trained on everyday photos, can be adapted to help scientists explore Earth’s hidden features from the deep ocean to beneath the surface. Our method uses minimal labeled data to detect over 20,000 new underwater volcanoes and also interprets complex subsurface structures in seismic data. By guiding the model with geological knowledge, we achieve accurate results across very different environments. This approach offers a scalable way to study Earth systems, even when data is sparse, and opens new paths for discovery in geoscience using foundation models.