FireRGBTNet

A Lightweight Forest Fire Detection Model Based on Efficient RGB-Thermal Fusion

Official PyTorch implementation of FireRGBTNet, published in Forests 2026, 17, 955.

UAV-based forest fire monitoring must detect very small targets under night-time illumination, dense smoke and complex terrain, while running on edge hardware. FireRGBTNet is an anchor-free RGB-T detector with 4.13 M parameters and 13.50 G FLOPs, reaching 95.9 % mAP@0.5 and 62.8 % mAP@0.5:0.95 on the RGBT-3M benchmark.

Architecture

Component	Module	File
Heterogeneous dual-stream backbone	RGBBlock, ThermalBlock	firergbtnet/models/backbone.py
Target-Enhanced Downsampling (with AWF)	TED	firergbtnet/models/backbone.py
Multi-Scale Semantic Alignment Enhancement	MSAE	firergbtnet/models/alignment.py
Multimodal Spatial Gated Fusion	MSGF (SGCA + MishGLU)	firergbtnet/models/fusion.py
Feature pyramid neck	FusionNeck	firergbtnet/models/neck.py
RT-DETR decoder with IoU-aware scoring	RTDETRDecoder	firergbtnet/models/head.py
Joint set-prediction loss (Hungarian + NWD)	SetCriterion	firergbtnet/loss.py

MSAE is an auxiliary supervision branch: it is active only during training and is dropped at inference, so it improves accuracy at zero deployment cost.

Module map

Where each piece of the paper lives, including the parts that carry no name in the publication and are therefore filed under conventional names:

Paper / original name	Code location	Note
RGBBlock, RGBUnit	models/backbone.py	standard-convolution ELAN block
ThermalBlock, ThermalUnit	models/backbone.py	large-kernel depthwise ELAN block
TED (Target-Enhanced Downsampling)	models/backbone.py	dual-branch downsample with AWF
MSAE	models/alignment.py	auxiliary alignment branch, Eqs. 12-18
MSGF, SGCA, MishGLU	models/fusion.py	spatial gated fusion, Eqs. 19-28
Feature pyramid neck	models/neck.py	top-down dual-stream FPN
RTDETRDecoder	models/head.py	RT-DETR head with deformable attention
Conv and the base units	models/basic.py	shared conv wrapper
FireRGBTNet (was RGBIRNet)	model.py	LightningModule, train/val loops
loss.py (was criterion.py)	loss.py	SetCriterion, HungarianMatcher, NWD
utils.py	boxes.py	box conversions, IoU, GIoU
metrics.py	metrics.py	mAP, P/R/F1, confusion matrix
data/datasets.py	data/dataset.py	RGBT-3M dataset and augmentation
train.py	tools/train.py	training entry point and callbacks

Installation

Requires Python 3.10+ and PyTorch 2.0+. The paper used PyTorch 2.4.1 + CUDA 11.8 on a Tesla V100.

git clone https://github.com/whynotnanfeng/FireRGBTNet.git
cd FireRGBTNet

conda create -n firergbtnet python=3.10 -y
conda activate firergbtnet

pip install -r requirements.txt

Dataset

Training and evaluation use RGBT-3M, a public UAV RGB-T forest fire dataset with 11,220 registered image pairs (7,854 train / 3,366 val) labelled for three classes: smoke, fire and person.

Zhang, Y.; Rui, X.; Song, W. A UAV-Based Multi-Scenario RGB-Thermal Dataset and Fusion Model for Enhanced Forest Fire Detection. Remote Sens. 2025, 17, 2593.

Download the dataset from the following mirrors:

• Quark Netdisk — access code hyx1
• Google Drive

Note that Quark Netdisk requires a mainland China phone number to access.

Organise the files in this YOLO-style layout, where each RGB image has a same-named thermal counterpart:

RGBT/
├── rgb/
│   ├── train/          # 7854 images
│   └── val/            # 3366 images
├── ir/
│   ├── train/
│   └── val/
└── labels/
    ├── train/          # <name>.txt  ->  cls cx cy w h   (normalised)
    └── val/

The dataset is not redistributed with this repository.

Training

python tools/train.py --data-root /path/to/RGBT --output-dir runs/firergbtnet

Defaults reproduce the configuration reported in the paper:

Parameter	Value
Input size	640 x 640
Epochs	250
Batch size	32
Optimizer	AdamW, lr 1e-3, weight decay 1e-4
LR schedule	2 % linear warm-up, then cosine annealing to 1e-6
EMA decay	0.9999
Mosaic / CutMix	disabled over the final 30 epochs
Object queries	300
Decoder	hidden dim 128, 2 layers

Useful overrides:

python tools/train.py --data-root /path/to/RGBT --epochs 300 --batch-size 16
python tools/train.py --data-root /path/to/RGBT --accelerator cpu --devices 1

Results

RGBT-3M validation split, 640 x 640 input, batch size 1 for throughput.

Model	P	R	F1	mAP@0.5	mAP@0.75	mAP@0.5:0.95	FPS	Params (M)	FLOPs (G)
Baseline	87.6	89.3	88.4	93.3	60.4	57.3	54.2	3.07	8.79
+ ThermalBlock	88.4	90.9	89.6	93.8	62.5	57.9	54.3	2.91	8.51
+ TED	89.7	90.2	90.0	94.1	62.1	57.8	46.1	2.87	7.88
+ MSGF	91.7	93.8	92.7	95.4	66.9	61.5	44.7	4.13	13.50
+ MSAE (FireRGBTNet)	91.8	94.7	93.2	95.9	68.9	62.8	44.7	4.13	13.50

Repository layout

FireRGBTNet/
├── firergbtnet/
│   ├── model.py           # FireRGBTNet LightningModule
│   ├── loss.py            # Hungarian matching, NWD, IoU-aware set loss
│   ├── metrics.py         # mAP, precision, recall, F1, confusion matrix
│   ├── boxes.py           # box conversions, IoU, GIoU
│   ├── models/
│   │   ├── basic.py       # Conv wrapper with BN folding for inference
│   │   ├── backbone.py    # RGBBlock, ThermalBlock, TED, dual-stream backbone
│   │   ├── alignment.py   # MSAE branch (CWA + cosine losses)
│   │   ├── fusion.py      # SGCA, MishGLU, MSGF
│   │   ├── neck.py        # top-down dual-stream FPN
│   │   └── head.py        # RT-DETR decoder, deformable attention
│   └── data/
│       └── dataset.py     # RGBT-3M dataset and paired augmentation
└── tools/
    └── train.py           # training entry point with EMA and close-mosaic callbacks

Notes

• TED implements the Adaptive Weighted Fusion of Eq. (10)-(11) with learnable branch weights, adding 16 learnable scalars in total.
• The neck operates on P3-P5; the stride-4 C2 stage is computed by the backbone but not consumed downstream.
• MishGLU uses a 9x9 depthwise convolution inside MSGF and a 3x3 one inside the RT-DETR decoder FFN.
• Parameter counts: 5.38 M for the full model during training, 4.27 M for deployment once MSAE is discarded.

Citation

@article{Ma_2026,
  title     = {FireRGBTNet: A Lightweight Forest Fire Detection Model Based on Efficient RGB–Thermal Fusion},
  author    = {Ma, Yifan and Shan, Weifeng and Wang, Maofa and Sui, Yanwei and Wang, Mengyu},
  journal   = {Forests},
  publisher = {MDPI AG},
  volume    = {17},
  number    = {8},
  pages     = {955},
  year      = {2026},
  month     = {Aug},
  issn      = {1999-4907},
  doi       = {10.3390/f17080955},
  url       = {https://doi.org/10.3390/f17080955}
}

Cite as: Ma, Y.; Shan, W.; Wang, M.; Sui, Y.; Wang, M. FireRGBTNet: A Lightweight Forest Fire Detection Model Based on Efficient RGB-Thermal Fusion. Forests 2026, 17(8), 955. https://doi.org/10.3390/f17080955

License

Code is released under the MIT License. The accompanying article is published by MDPI under CC BY 4.0.