---
title: CEA List FrugalAI Challenge
emoji: 🔥
colorFrom: red
colorTo: yellow
sdk: docker
pinned: false
license: apache-2.0
short_description: YOLO for low-emission Early Fire Detection
---


# YOLO for Early Fire Detection

## Team ([CEA List, LVA](https://kalisteo.cea.fr/index.php/ai/))
- Renato Sortino
- Aboubacar Tuo
- Charles Villard
- Nicolas Allezard
- Nicolas Granger
- Angélique Loesch
- Quoc-Cuong Pham

## Model Description

YOLO model for early fire detection in forests, proposed as a solution for the [Frugal AI Challenge 2025](https://frugalaichallenge.org/), image task. 

## Training Data

The model uses the following datasets:
| Dataset | Number of samples | Number of instances | 
|----------|----------|----------|
| [pyronear/pyro-sdis](https://huggingface.co/datasets/pyronear/pyro-sdis) | 29,537 | 28,167 |
| [D-Fire](https://github.com/gaiasd/DFireDataset) | 10,525 | 11,865 |
| [Wildfire Smoke Dataset](https://www.kaggle.com/datasets/gloryvu/wildfire-smoke-detection/data) | ~12,300 | 11,539  |
| [Hard Negatives](https://github.com/aiformankind/wildfire-smoke-dataset) | ~5,000 | ~5,000 |
| Synthetic Dataset | ~5,000 | ~5,000 |

## Performance

### Model Architecture
The model is a YOLO-based object detection model, that does not depend on NMS in inference. 
Bypassing this operation allows for further optimization at inference time via tensor decomposition.

### Metrics
| Model | Accuracy | Precision | Recall | meanIoU | Wh | gCO2eq
|----------|----------|----------|----------|----------|----------|----------|
| YOLOv10s | 0.87 | 0.88 | 0.98 | 0.84 | 6.77 | 0.94 |
| YOLOv10m | 0.88 | 0.87 | 0.99 | 0.88 | 8.39 | 1.16 |
| YOLOv10m + Spatial-SVD | 0.85 | 0.86 | 0.97 | 0.82 | 8.24 | 1.14 |

Environmental impact is tracked using [CodeCarbon](https://codecarbon.io/), measuring:
- Carbon emissions during inference (gCO2eq)
- Energy consumption during inference (Wh)

This tracking helps establish a baseline for the environmental impact of model deployment and inference.

## Limitations and future work
- It may fail to generalize to night scenes or foggy settings
- It is subject to false detections, especially at low confidence thresholds
- Clouds at ground level can be misinterpreted as smoke
- It would be better to use temporal-aware models trained on videos

```