Gemma Earth (Gemma 3 4B IT LoRA for Earth Observation)

Model Description

Gemma Earth is a domain-adapted Earth Observation model built by fine-tuning Google Gemma 3 4B IT with LoRA adapters for satellite scene understanding.

The current release focuses on multi-label land-use and land-cover classification from the EarthDial BigEarthNet subset, with a pipeline designed to extend to additional EarthDial tasks.

This project provides an end-to-end JAX stack pipeline for dataset preparation, LoRA fine-tuning, checkpointing, evaluation, and inference serving.

Project repository: https://github.com/haruiz/gemma_earth

Base model: google/gemma-3-4b-it
Adaptation: LoRA
Training stack: JAX + Flax (NNX) + Tunix + Qwix + Optax + Orbax + Grain
Primary task: remote-sensing scene classification (multi-label)

Intended Use

Primary Use

Earth Observation and remote-sensing research
Multi-label scene classification on EarthDial/BigEarthNet-style samples
Benchmarking and experimentation with TPU-based JAX fine-tuning workflows

Training Data

Dataset source: akshaydudhane/EarthDial-Dataset
Current focus: EarthDial classification subset (BigEarthNet)
Training setup in this project uses sampled subsets and validation splits configured via environment variables

Training Procedure

Hardware: Google Cloud TPU v5litepod-8
LoRA configuration (default project setup):
- Rank: 32
- Alpha: 64.0
Typical sequence length: 768
Optimizer schedule includes warmup + decay (Optax)

Evaluation Summary

Benchmark run size: 1500 samples

Metric	Baseline	Fine-tuned	Delta (absolute)
Exact Match	2.53%	22.80%	+20.27 pp
Sample Precision	27.27%	68.53%	+41.27 pp
Sample Recall	10.35%	71.86%	+61.51 pp
Sample F1	14.18%	68.16%	+53.98 pp
Sample Jaccard	10.35%	57.47%	+47.12 pp
Micro Precision	29.57%	63.56%	+33.99 pp
Micro Recall	9.25%	67.00%	+57.75 pp
Micro F1	14.09%	65.24%	+51.15 pp
Macro F1	6.53%	31.50%	+24.97 pp

The adapted model significantly improves over baseline across all reported metrics.

How To Use

This model is intended to be used with the Gemma Earth codebase.

# Example: run one-image evaluation using a Hugging Face checkpoint directory
python scripts/one_example_eval.py \
  --model-checkpoint-source huggingface \
  --model-dir /path/to/hf_checkpoint_dir \
  --image-path /path/to/image.jpg

# Example: run benchmark evaluation with Hugging Face checkpoint source
python main.py benchmark \
  --num-examples 1500 \
  --eval-restore-policy permissive \
  --model-checkpoint-source huggingface

Citation

If you use this model, please cite:

@misc{gemma_earth_2026,
  title={Gemma Earth: Fine-tuning Gemma for Remote Sensing Scene Classification},
  author={Henry Ruiz},
  year={2026},
  howpublished={GitHub repository},
  url={https://github.com/haruiz/gemma_earth},
}

Related references:

@misc{soni2024earthdial,
  title={EarthDial: Turning Multi-sensory Earth Observations to Interactive Dialogues},
  author={Soni, Sagar and Dudhane, Akshay and Debary, Hiyam and Fiaz, Mustansar and Munir, Muhammad Akhtar and Danish, Muhammad Sohail and Fraccaro, Paolo and Watson, Campbell D and others},
  year={2024},
  eprint={2412.15190},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  doi={10.48550/arXiv.2412.15190},
  url={https://arxiv.org/abs/2412.15190}
}

@misc{sumbul2019bigearthnet,
  title={BigEarthNet: A Large-Scale Benchmark Archive For Remote Sensing Image Understanding},
  author={Sumbul, Gencer and Charfuelan, Marcela and Demir, Beg{"u}m and Markl, Volker},
  year={2019},
  eprint={1902.06148},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  doi={10.48550/arXiv.1902.06148},
  url={https://arxiv.org/abs/1902.06148}
}