Earth observation has no shortage of data. What it still lacks are practical, scalable workflows for adapting modern multimodal foundation models to geospatial tasks without turning experimentation and training into a months-long infrastructure effort. In this talk, we introduce GemmaEarth, an open-source post-training and benchmarking framework designed to adapt Google’s Gemma 3 4B IT model for Earth Observation (EO) understanding using a TPU-native JAX ecosystem.

GemmaEarth leverages tools such as Tunix, Grain, Optax, Orbax, and Qwix to provide a scalable and reproducible workflow for parameter-efficient fine-tuning on Google Cloud TPU v5litepod-8. Using the EarthDial dataset as an initial benchmark, the framework focuses on multi-label satellite scene classification while establishing a flexible foundation for broader EO applications, including multimodal reasoning, scene understanding, and geospatial analysis.

The session will present the end-to-end workflow behind GemmaEarth, covering dataset preparation, LoRA-based post-training, distributed TPU experimentation, evaluation, benchmarking, and deployment considerations. Attendees will gain practical insights into adapting open multimodal models for remote sensing tasks with the JAX ecosystem and Google Cloud TPUs, as well as lessons learned from building reproducible and scalable EO-focused AI pipelines for both research and production environments.