haruiz/gemmaearth
Gemma Earth is a 4.3 billion parameter domain-adapted Earth Observation model developed by haruiz, fine-tuned from Google's Gemma 3 4B IT using LoRA adapters. It specializes in multi-label land-use and land-cover classification from satellite imagery, particularly on the EarthDial BigEarthNet subset. This model significantly improves classification metrics over its baseline, making it suitable for remote-sensing research and benchmarking JAX fine-tuning workflows.
Loading preview...
Gemma Earth: Domain-Adapted for Earth Observation
Gemma Earth is a specialized model developed by haruiz, built upon Google's Gemma 3 4B IT architecture. It leverages LoRA (Low-Rank Adaptation) fine-tuning to excel in Earth Observation tasks, specifically focusing on satellite scene understanding. The model was trained using a JAX-based stack on Google Cloud TPU v5litepod-8 hardware.
Key Capabilities
- Multi-label Scene Classification: Primarily designed for land-use and land-cover classification from satellite imagery, utilizing the EarthDial BigEarthNet subset.
- Significant Performance Improvement: Achieves substantial gains over the baseline model across various metrics, including a +20.27 pp increase in Exact Match and a +53.98 pp increase in Sample F1 on a 1500-sample benchmark.
- JAX/Flax Integration: Developed with an end-to-end JAX stack pipeline for dataset preparation, LoRA fine-tuning, checkpointing, evaluation, and inference serving.
Intended Use Cases
- Earth Observation Research: Ideal for academic and research applications in remote sensing.
- Benchmarking: Useful for evaluating and experimenting with TPU-based JAX fine-tuning workflows.
- BigEarthNet Analysis: Specifically adapted for multi-label scene classification on EarthDial/BigEarthNet-style datasets.
This model is intended to be used with the associated Gemma Earth codebase for optimal performance and integration.