---
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen2.5-VL-3B-Instruct
tags:
- remote-sensing
- earth-observation
- geospatial
- qwen2_5_vl
- grpo
---

# GeoVLM_R1_Classification

> **Disclaimer**
>
> THIS MODEL IS PROVIDED “AS IS,” WITHOUT WARRANTIES OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, ACCURACY, OR NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR THEIR AFFILIATES BE LIABLE FOR ANY DAMAGES ARISING FROM ACCESS TO OR USE OF THIS MODEL.

This model is released **without a license**. It is not open-source software.

## Model description

`GeoVLM_R1_Classification` is the **Scene classification** checkpoint of **GeoVLM-R1**, from the paper
[GeoVLM-R1: Reinforcement Fine-Tuning for Improved Remote Sensing Reasoning](https://arxiv.org/abs/2509.25026).

- **Architecture:** `Qwen2_5_VLForConditionalGeneration` (3B parameters, bfloat16)
- **Training:** Qwen2.5-VL-3B-Instruct, then supervised fine-tuning on EarthDial-Instruct, then
  GRPO reinforcement fine-tuning with task-aware rewards on `GRPO_BigEarth_classification.jsonl`
- **Task:** Scene classification
- **Benchmarks:** AID, UCMerced, WHU-RS19, BigEarthNet, fMoW

GeoVLM-R1 has one checkpoint per task. All 9 are in
[ibm-research/GeoVLM-R1](https://huggingface.co/ibm-research/GeoVLM-R1).

## Usage

```python
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

model_path = "path/to/GeoVLM_R1_Classification"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path, torch_dtype=torch.bfloat16, device_map="auto").eval()
processor = AutoProcessor.from_pretrained(model_path)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": "example.png", "resized_height": 448, "resized_width": 448},
        {"type": "text", "text": "Describe this image."},
    ],
}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
images, videos = process_vision_info(messages)
inputs = processor(text=[text], images=images, videos=videos,
                   padding=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128)
print(processor.batch_decode([o[len(i):] for i, o in zip(inputs.input_ids, out)],
                             skip_special_tokens=True)[0])
```

To run the benchmark, use `GeoVLM_inference/generate_inference_classification.py` from the
[GeoVLM-R1 Toolkit](https://github.com/IBM/GeoVLM-R1-Toolkit) with `--model_path path/to/GeoVLM_R1_Classification`.

## Citation

```bibtex
@article{fiaz2025geovlmr1,
  title   = {GeoVLM-R1: Reinforcement Fine-Tuning for Improved Remote Sensing Reasoning},
  author  = {Fiaz, Mustansar and Debary, Hiyam and Fraccaro, Paolo and Paudel, Danda and
             Van Gool, Luc and Khan, Fahad and Khan, Salman},
  journal = {arXiv preprint arXiv:2509.25026},
  year    = {2025}
}
```

## Contact

Mustansar Fiaz <mustansar.fiaz@ibm.com>