Φ-talk - Training and deploying pre trained models for remote sensing data скачать в хорошем качестве

Φ-talk - Training and deploying pre trained models for remote sensing data

In this Φ-talk, Gabriel Tseng, Research Scientist at the Allen Institute for Artificial Intelligence (Ai2), presents ongoing research on self-supervised learning for satellite data and the challenges of deploying these models at scale. Machine learning for satellite data has wide-ranging societal applications, from crop monitoring to climate modelling, but labelled training data is often scarce or unavailable. This talk explores how the structure of remote sensing data itself, including its temporal dimension and multi-sensor availability, can be leveraged to train performant models without relying on large labelled datasets. Gabriel also discusses the practical challenges of making these models accessible to the broader remote sensing community, drawing on his experience at Ai2's OlmoEarth team and previous work with NASA Harvest on global cropland and crop type mapping. Topics covered: Self-supervised learning for remote sensing, pre-training strategies for satellite data, temporal and multi-sensor data, model deployment at scale, and Earth Observation foundation models. About the speaker: Gabriel Tseng is a research scientist at Ai2, working on the OlmoEarth platform. He holds a PhD from McGill University / Mila, where he investigated pre-training algorithms for remote sensing models under David Rolnick and Hannah Kerner. About the ESA Φ-Lab Collaborative Innovation Network The CIN is a community hosted by ESA's Φ-Lab, bringing together researchers, engineers and innovators working at the frontier of Earth Observation and AI. 🌍 ESA Φ-Lab: https://philab.esa.int 🔗 CIN LinkedIn community: https://bit.ly/4luuNDH