SAINet - Latest smoke detection checkpoint for the SAI

This repository always publishes the latest production checkpoint of the computer vision model used in the Sistema de Alerta de Incendios (SAI).

The repo is intentionally minimal: it exposes a single checkpoint file (model/SAINet_v11.1.pt) plus a compact configuration file so that others can download, run and reproduce the model under the terms of the license.

Current checkpoint

• Model name: SAINet v11.1
• Architecture: Ultralytics YOLO26m (medium)
• Checkpoint path: model/SAINet_v11.1.pt
• Training hyperparameters: model/train_hyperparams.yaml
• Main training dataset: SAINetset v8.0 (~65K images, smoke & fire detection)
• Evaluation datasets:
  • Validation split of SAINetset 8.0 (generalization)
  • In-situ SAI datasets (La Rancherita and La Serranita nodes, Cordoba, Argentina)

---

Repository layout

The repository is kept deliberately small:

• LICENSE
  The GNU Affero General Public License v3.0 (AGPL-3.0) that applies to this model.

• README.md
  This file. Explains the purpose of the repo, datasets, benchmarks and how to use the current checkpoint.

• model/train_hyperparams.yaml
  YOLO training hyperparameters for the current checkpoint (task, epochs, batch size, image size, optimizer, data augmentations, etc.). Use this file as the single source of truth for the training configuration.

• model/SAINet_v11.1.pt
  The only checkpoint published here. It always corresponds to the latest production model used in the SAI system.

Previous checkpoints are not stored as extra files in this repository.

---

Quick usage

The checkpoint is an Ultralytics YOLO26 model trained for two detection classes: smoke and fire.

Python example

from ultralytics import YOLO

# Load the latest SAINet checkpoint from this repo
model = YOLO("model/SAINet_v11.1.pt")

# Run inference on an image (both smoke and fire)
results = model("path/to/an/image.jpg", conf=0.25)
results.show()  # or results.save()

# Run inference for smoke only
results = model("path/to/an/image.jpg", conf=0.25, classes=[0])

Inference notes

• The model is designed for early wildfire smoke in outdoor landscapes.
• The operational policy of the SAI is recall-first (missing a plume is worse than raising an extra alert), so we typically:
  • use low confidence thresholds,
  • calibrate IoU / NMS settings, and
  • rely on temporal consensus at system level to control false positives.

When you deploy this model in a different context, you should:

• Re-tune conf, iou and NMS options for your scenario.
• Always check performance on:
  • a held-out test set, and
  • a real-world evaluation set that matches your deployment conditions.

---

Training setup

Training is done with Ultralytics YOLO26 on top of PyTorch, starting from a COCO-pretrained YOLO26m (medium) backbone and neck.

All relevant training settings are stored in:

• model/train_hyperparams.yaml

This YAML file is meant to be the canonical training config for this checkpoint and includes (at least):

• task and model type (e.g. task: detect, model: yolo26m),
• list of classes used for training (smoke and fire),
• number of epochs, batch size, image size,
• optimizer and learning rate schedule.

---

Datasets

Current SAINet training dataset (SAINetset 8.0)

Download the dataset: SAINetset v8.0 on Hugging Face

The main training dataset, SAINetset 8.0, is a curated smoke & fire detection dataset designed for outdoor, long-range wildfire detection. It combines multiple data sources:

1. D-Fire - a drone-captured fire & smoke dataset with realistic forest fire scenes.
2. Pyronear (pyro-sdis) - wildfire detection images from the Pyronear project.
3. SAI field data - real-world images from deployed SAI nodes in Córdoba, Argentina.

SAINetset 8.0 is defined at the level of images:

• Positive image: at least one smoke or fire bounding box.
• Negative image: no smoke/fire annotations (background only).

Composition by split (images)

Split	Total images	Positives	Negatives	Positive %	Negative %
Train	56,815	38,113	18,702	67.1 %	32.9 %
Val	7,894	4,828	3,066	61.2 %	38.8 %
Overall	64,709	42,941	21,768	66.4 %	33.6 %

Class distribution (bounding boxes)

Class	Train boxes	Val boxes	Total boxes
smoke	41,806	5,478	47,284
fire	13,146	1,458	14,604

The dataset is intentionally rich in:

• diverse positive smoke and fire patterns, and
• hard negatives that are typical sources of false alarms in real SAI deployments (clouds, haze, reflections, steam, etc.).

Download SAINetset

The dataset is publicly available on Hugging Face. You can download it using any of these methods:

Option 1: Git LFS (recommended)

git lfs install
git clone https://huggingface.co/datasets/SAINetset/SAINetset_v8.0

Option 2: Hugging Face Datasets library

from datasets import load_dataset

ds = load_dataset("SAINetset/SAINetset_v8.0")
sample = ds["train"][0]

Option 3: Direct use with Ultralytics YOLO

from ultralytics import YOLO

model = YOLO("yolo26m.pt")
model.train(data="path/to/sainetset/data.yaml", epochs=100)

In-situ evaluation datasets

To evaluate performance under real operating conditions, the model is also tested on separate in-situ datasets from production SAI nodes:

• La Rancherita (Cordoba, Argentina) - first deployed node.
• La Serranita (Cordoba, Argentina) - second deployed node.

These datasets contain:

• a small number of confirmed smoke events (positives), and
• a large number of typical background frames (hills, vegetation, clouds, changing lighting).
• No synthetic data and no images from other regions.
• These datasets are used only for evaluation, not for training.

---

Benchmarks (current checkpoint)

This section summarizes the performance of the current checkpoint on:

1. The SAINetset 8.0 test split (generalization on the main dataset).
2. The in-situ evaluation dataset (La Rancherita node).

All metrics are for smoke detection only (no fire class) using standard Ultralytics metrics:

• P = precision
• R = recall
• F1 = harmonic mean of P and R
• mAP@50 = mean Average Precision at IoU 0.5
• mAP@50-95 = mean AP averaged over IoU thresholds from 0.5 to 0.95

Two inference configurations are reported:

• Recall-oriented - tuned to minimize false negatives (FN ~ 0), accepting more false positives.
• Precision-oriented - tuned to reduce false positives while maintaining acceptable recall.

1. SAINetset 8.0 - test split

Config	P	R	F1	mAP@50	mAP@50-95
Recall-oriented (conf=0.15, IoU=0.10)	0.833	0.797	0.815	0.842	0.494
Precision-oriented (conf=0.15, IoU=0.40)	0.854	0.774	0.812	0.844	0.491

2. In-situ dataset - La Rancherita

Config	P	R	F1	mAP@50	mAP@50-95
Recall-oriented (conf=0.25, IoU=0.70)	0.859	0.861	0.860	0.906	0.697
Precision-oriented (conf=0.25, IoU=0.10)	0.883	0.842	0.862	0.900	0.697

---

Related Links

• SAI Project Website: sainet.info
• Training Dataset: SAINetset v8.0 on Hugging Face
• Model Repository: sai-sainet-latest on GitHub

---

License

This model and all files in this repository are released under the GNU Affero General Public License v3.0 (AGPL-3.0).

If you use this model to provide a service to third parties over a network, the AGPL-3.0 requires you to make the corresponding source code available, including any local modifications, in accordance with the license terms.