Abstract

FLAME2-DT (Forest Fire Detection Dataset with Dual-modality Labels) is a comprehensive multi-modal dataset specifically designed for UAV-based forest fire detection research. The dataset consists of 1,280 paired RGB-thermal infrared images captured by a Mavic 2 Enterprise Advanced UAV system, with high-resolution (640×512) and precise pixel-level annotations for both fire and smoke regions. This dataset addresses critical challenges in forest fire detection by providing paired multi-modal data that captures the complementary characteristics of visible light and thermal imaging. The RGB images contain 2,496 fire bounding boxes and 4,404 smoke bounding boxes, while the thermal infrared images include 27,117 fire bounding boxes. Statistical analysis reveals distinct scale and distribution patterns: approximately 80% of fire regions occupy less than 5% of the image area with discrete distribution, while over 60% of smoke regions cover more than 12% with continuous patterns. FLAME2-DT is organized into five specialized packages to facilitate different research scenarios: original dataset, RGB-specific, thermal IR-specific, RGB with dual-modality labels, and complete fusion packages. The dataset is split into training (80%) and validation (20%) sets, providing a standardized benchmark for evaluating multi-modal forest fire detection algorithms. This dataset contributes to the advancement of forest fire detection research by: 1. Providing precisely registered multi-modal image pairs 2. Offering comprehensive pixel-level annotations verified through multi-expert cross-validation 3. Supporting the development of lightweight, real-time detection systems for UAV applications 4. Enabling comparative analysis of single-modal and multi-modal detection approaches

Instructions:

Quick Links

Download:

- Baidu Netdisk: https://pan.baidu.com/s/1bTm04RSyCaLC5TlmZG6SkQ?pwd=qfds

- Extraction Code: qfds

1. Introduction

FLAME2-DT is an enhanced version of the FLAME2 dataset, specifically designed for multi-modal forest fire detection research. The dataset features paired RGB-thermal infrared images captured by a Mavic 2 Enterprise Advanced UAV system, with comprehensive pixel-level annotations for both fire and smoke regions. This dataset aims to address the limitations in the original FLAME2 dataset, including missing annotations, resolution constraints, and spatial registration biases.

2. Dataset Characteristics

Target Distribution Analysis

Key findings from statistical analysis:

1. Scale Characteristics

Fire: ~80% of bounding boxes occupy <5% of image area (small-scale, discrete distribution)
Smoke: >60% of bounding boxes occupy >12% of image area (large-scale, continuous distribution)

2. Spatial Distribution

Fire: Multi-centered, scattered pattern
Smoke: Relatively dispersed distribution of center points

3. Dataset Development

Data Collection

Equipment: Mavic 2 Enterprise Advanced UAV dual-camera system
Resolution: 640×512 pixels (both RGB and thermal IR)
Processing: Feature point matching and spatial transformation consistency analysis

Annotation Protocol

RGB Images:

- 2,496 fire bounding boxes

- 4,404 smoke bounding boxes

Thermal IR Images:

- 27,117 fire bounding boxes

- No smoke annotations (due to thermal imaging limitations)

4. Dataset Organization

The dataset contains five specialized packages:

1. origin_dataset

Complete frame-extracted original images
Full annotation files

2. RGB+RGB_labels

RGB modality images
RGB-specific annotations
Training/validation split

3. IR+IR_labels

Thermal infrared images
IR-specific annotations
Training/validation split

4. RGB+RGB_IR_labels

RGB images
Combined RGB-IR annotations
Multi-modal analysis optimization

5. RGB_IR_fuse_labels

Complete RGB-IR image pairs
Comprehensive annotation set
Full multi-modal capability

5. Dataset Scale

Total Image Pairs: 1,280
Training Set: 1,024 pairs (80%)
Validation Set: 256 pairs (20%)

6. Applications

Primary Research Areas

UAV-based forest fire detection
Multi-modal object detection
Remote sensing image analysis
Emergency response systems

Technical Development

Multi-modal fusion algorithms
Lightweight deep learning models
Real-time detection systems

7. Citation

If you use the FLAME2-DT dataset in your research, please cite our paper:

......

Dataset Files

Files have not been uploaded for this dataset

Datasets

Standard Dataset

FLAME2-DT