---
base_model: 5ch4um1/lfm2.5-vrsbench-lora-450m
tags:
- vision-language
- satellite
- methane
- sentinel-2
- methane-detection
- bounding-box
- lfm2
- gguf
- vrsbench
---

# LFM2.5-VL-450M VRSBench + MethaneS2CM Expert

## Model Description

This is a fine-tuned version of LiquidAI's LFM2.5-VL-450M vision-language model, specialized for methane plume detection in satellite imagery. The model was trained in two stages:

1. **VRSBench Training**: Base training on VRSBench dataset
2. **MethaneS2CM Fine-tuning**: Additional training on MethaneS2CM dataset for methane detection

The model can detect methane plumes in Sentinel-2 satellite images and provide bounding box coordinates for the plume location.

## Training Details

### Stage 1: VRSBench Pre-training
- **Base Model**: LFM2.5-VL-450M
- **Dataset**: VRSBench (Vision Reasoning and Scene Understanding Benchmark)
- **Epochs**: 1
- **Method**: LoRA (r=16, alpha=32)

### Stage 2: Methane Detection Fine-tuning
- **Base Model**: VRSBench-trained model (`5ch4um1/lfm2.5-vrsbench-lora-450m`)
- **Dataset**: **MethaneS2CM** (Methane Sentinel-2 Community Model)
  - Source: [H1deaki/MethaneS2CM on Hugging Face](https://huggingface.co/datasets/H1deaki/MethaneS2CM)
  - 257,096 training samples, 60,567 test samples
  - Hand-annotated plume masks for bounding box extraction
  - Sentinel-2 bands: SWIR22 (Band 12) → Red, NIR08 (Band 8) → Green, Red (Band 4) → Blue
- **Training Samples**: 20,000 (10k positive with plumes, 10k negative)
- **Epochs**: 2
- **Method**: LoRA (r=16, alpha=32)
- **Hardware**: Local GPU training (no Ray/distributed)

## Evaluation Results

### Methane Detection (500 test samples)

| Metric | BASE VRSBench | METHANE EXPERT (this model) | Improvement |
|--------|-------------------|---------------------------|--------------|
| Accuracy | 49.80% | 51.00% | +1.20% |
| Precision | 0.00% | 50.79% | +50.79% |
| Recall | 0.00% | 88.89% | +88.89% |
| F1 Score | 0.00% | 64.65% | +64.65% |
| Mean IoU (bbox) | 0.0000 | 0.5879 | +0.5879 |

**Key Results:**
- The base VRSBench model cannot predict methane plumes or bounding boxes (0% recall, 0 IoU)
- The Methane Expert model achieves **88.89% recall** - excellent at detecting plumes
- Bounding box predictions have **0.5879 mean IoU** (good localization)
- Precision is 50.79% - the model tends to over-predict plumes (217 false positives vs 31 true negatives)
- 113 out of 224 predicted bounding boxes have IoU > 0.5 (good localization)

## Usage

### With llama.cpp

```bash
# Download Q4_K_M quantized version (recommended)
wget https://huggingface.co/5ch4um1/lfm2.5-vrsbench-MethaneS2CM-methane-lora-450m/resolve/main/lfm2.5-vrsbench-methane-450m-q4_k_m.gguf

# Run inference
./llama-cli -m lfm2.5-vrsbench-methane-450m-q4_k_m.gguf \
  --image satellite_methane_image.png \
  -p "Is there a methane plume in this satellite image? If yes, provide the bounding box as [x1,y1,x2,y2] normalized to 0-1."
```

### With Transformers

```python
from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image
import json

model = AutoModelForVision2Seq.from_pretrained(
    "5ch4um1/lfm2.5-vrsbench-MethaneS2CM-methane-lora-450m",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("5ch4um1/lfm2.5-vrsbench-MethaneS2CM-methane-lora-450m")

image = Image.open("satellite_methane_image.png")
prompt = "Is there a methane plume in this satellite image? If yes, provide the bounding box as [x1,y1,x2,y2] normalized to 0-1."

conversation = [
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": prompt}
    ]}
]

text = processor.apply_chat_template(conversation, add_generation_prompt=True)
inputs = processor(text=text, images=image, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=100)
pred_text = processor.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(json.loads(pred_text))
```

## GGUF Quantizations

| Version | Size | Description |
|---------|------|-------------|
| F16 | 679 MB | Full precision (16-bit) |
| Q8_0 | 362 MB | 8-bit quantization |
| Q4_K_M | 219 MB | 4-bit quantization (recommended for most use cases) |

## Training Dataset: MethaneS2CM

The model was trained on the **MethaneS2CM** (Methane Sentinel-2 Community Model) dataset:

- **Dataset**: [H1deaki/MethaneS2CM](https://huggingface.co/datasets/H1deaki/MethaneS2CM)
- **Paper**: "MethaneS2CM: A Dataset for Multispectral Deep Methane Emission Detection" (Liu et al., 2025)
- **Data**: 257k+ samples from Sentinel-2 imagery (2016-2024)
- **Annotations**: Hand-annotated plume masks for bounding box extraction
- **Bands used**: False color composite (SWIR22→R, NIR08→G, Red→B)
- **Task**: Methane plume detection with bounding box regression

## Model Sources

- **Base Model**: [LiquidAI/LFM2.5-VL-450M](https://huggingface.co/LiquidAI/LFM2.5-VL-450M)
- **VRSBench Model**: [5ch4um1/lfm2.5-vrsbench-lora-450m](https://huggingface.co/5ch4um1/lfm2.5-vrsbench-lora-450m)
- **Methane Expert Model**: [5ch4um1/lfm2.5-vrsbench-MethaneS2CM-methane-lora-450m](https://huggingface.co/5ch4um1/lfm2.5-vrsbench-MethaneS2CM-methane-lora-450m)
- **MethaneS2CM Dataset**: [H1deaki/MethaneS2CM](https://huggingface.co/datasets/H1deaki/MethaneS2CM)

## Limitations

- Model trained on 32x32 Sentinel-2 patches - may not generalize to other resolutions
- Performance depends on quality of false color composite (SWIR22/NIR08/Red)
- Hand-annotated masks may contain annotation errors
- Best performance on methane plumes similar to training distribution

## Training Environment

- **Framework**: Transformers + PEFT (LoRA) + leap-finetune
- **Hardware**: Local GPU (CUDA)
- **Training Scripts**: Available in the [cookbook repository](https://github.com/anomalyco/opencode)