---
base_model: LiquidAI/LFM2.5-VL-450M
datasets:
  - munish0838/crop-burn-detection-labeled
tags:
  - satellite
  - remote-sensing
  - sentinel-2
  - stubble-burning
  - fine-tuned
  - vision-language-model
license: cc-by-4.0
---

# crop-burn-detector-v2

Fine-tuned **LFM2.5-VL-450M** (Liquid AI) on the [crop-burn-detection-labeled](https://huggingface.co/datasets/munish0838/crop-burn-detection-labeled) dataset.

Given a pair of Sentinel-2 satellite images (RGB + SWIR), the model outputs a structured JSON burn assessment for a 5 km × 5 km agricultural tile.

## Inference

```python
from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image

model = AutoModelForImageTextToText.from_pretrained("munish0838/crop-burn-detector-v2", torch_dtype="bfloat16", device_map="auto")
processor = AutoProcessor.from_pretrained("munish0838/crop-burn-detector-v2")

rgb_image  = Image.open("tile_rgb.png")
swir_image = Image.open("tile_swir.png")

messages = [
    {"role": "system", "content": "You are an expert in analyzing Sentinel-2 satellite imagery for crop residue burning detection in northern India."},
    {"role": "user", "content": [
        {"type": "image", "image": rgb_image},
        {"type": "image", "image": swir_image},
        {"type": "text",  "text": "Analyze this RGB + SWIR satellite tile and return a JSON burn assessment."},
    ]},
]

inputs = processor.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.batch_decode(out, skip_special_tokens=True)[0])
```

## Output Schema

```json
{
  "burn_detected": true,
  "burn_severity": "moderate",
  "burn_fraction_estimate": 0.25,
  "burn_freshness": "recent",
  "active_smoke_visible": false,
  "vegetation_phase": "post_harvest",
  "image_quality_limited": false,
  "notes": "Dark brownish-red burn scars with rectangular field boundaries visible in SWIR."
}
```

## Training Details

| | |
|---|---|
| Base model | LiquidAI/LFM2.5-VL-450M |
| Training data | munish0838/crop-burn-detection-labeled |
| Train samples | 1,098 |
| Epochs | 5 |
| Learning rate | 0.0001 |
| LoRA rank | 32 |
| Effective batch | 16 |