Imaging Science & Engineering seminar | Nathan Jacobs

About this event

Mapping What Satellites Can’t See

Satellites image the entire planet every few days, yet many of the things we most want to map are invisible from above: what a place sounds like, which species live there, when plants flower, how people experience it. This talk describes recent work from my lab on learning such maps by pairing overhead imagery with sparse, noisy ground-level observations, including geotagged photos, audio recordings, and citizen-science species reports. I’ll focus on ideas that apply well beyond remote sensing: learning a shared embedding space across many modalities when most samples are missing most modalities, modeling uncertainty explicitly, imposing known hierarchical structure on learned representations, and using implicit neural representations to encode a continuous, spatially adaptive field over the globe.