FLAME Dataset: Wildfire Imaging & Segmentation

• FLAME dataset is a family of high-resolution, multimodal UAV imagery datasets designed for wildfire detection and segmentation, with radiometric thermal calibration for per-pixel temperature mapping.
• It employs advanced imaging protocols with synchronized RGB and thermal modalities, detailed annotations, and comprehensive calibration for both classification and segmentation tasks.
• FLAME 3 enhances previous versions by incorporating radiometric thermal TIFFs, improved calibration, and lightweight, real-time inference capabilities for UAV-based wildfire surveillance.

FLAME refers to a family of datasets primarily designed either for remote sensing wildfire surveillance, federated learning in robotics, language modeling for spreadsheet formulas, or evaluation of financial LLMs. This article focuses on the FLAME datasets as developed for wildfire imagery and segmentation, as these constitute the most widely adopted and technically mature reference under the FLAME acronym. The FLAME dataset lineage provides high-resolution, multimodal imagery targeted at machine learning for wildfire detection, segmentation, and modeling. Recent evolutions (notably FLAME 3) include radiometric thermal imagery, enabling ground-truth per-pixel temperature recovery from UAV-acquired (Unmanned Aerial Vehicle) datasets in prescribed and wildland fire scenarios (Shamsoshoara et al., 2020, Hopkins et al., 2024, Zhao et al., 2024).

1. Dataset Series and Scope

The FLAME dataset series comprises three major generations: FLAME 1, FLAME 2, and FLAME 3.

• FLAME 1 (2020): Acquired during a single prescribed pile burn, capturing co-registered visible (RGB) and thermal (non-radiometric paletted) images. Total dataset size: 47,992 images, roughly balanced between “fire” and “no fire” classes. Data supported binary image classification and basic segmentation tasks (Shamsoshoara et al., 2020).
• FLAME 2 (2022): Expanded to six prescribed burns, acquisition included paired RGB and thermal (again non-radiometric) images, supporting two-stream deep learning studies. Supplementary data: pre-/post-burn orthomosaics, weather logs, and pointclouds. 53,451 RGB–IR image pairs collected, with segmentation masks on ~2,003 images.
• FLAME 3 (2024): First dataset to provide side-by-side radiometric thermal Tag Image File Format (TIFF) images and high-res RGB. Six prescribed burns across diverse ecosystems with variable fuel types, fuel moistures (4–15 %), and burn behaviors. 13,997 oblique images, plus 800–1,300 time-series (“nadir”) frames per modeling subset, all with per-pixel temperature ground truth. Image resolution: thermal 640×512 px (TIFF), RGB up to 8,000×6,000 px (Hopkins et al., 2024).

Table 1 summarizes FLAME lineages and technical advances.

FLAME Version	Year	Sensor Modalities	Main Novelty
FLAME 1	2020	RGB, paletted thermal	Pile burns, non-radiometric
FLAME 2	2022	RGB, paired paletted IR	Multi-burn, two-stream pairing
FLAME 3	2024	RGB, radiometric thermal	Per-pixel temp, nadir series

2. Data Acquisition Methodology and Calibration

Each dataset stage adopted detailed protocols for multimodal imaging:

• Flight Planning: Includes pre-burn mapping (SfM at 122 m AGL), oblique imaging (80% overlap), ground control plates (RTK GPS), “no-fire” negative sampling, active-burn passes (flaming front focus), and specialized nadir thermal “plots” (hover, overhead, 3–5 s timesteps) (Hopkins et al., 2024).
• Sensor Suite: Platforms include Autel EVO II Dual 640 T Enterprise, DJI Mavic 2 Enterprise Advanced, DJI Matrice 30T, and DJI Matrice 200 with Zenmuse X4S. All relevant cameras provide synchronized RGB and either paletted or radiometric longwave thermal data.
• Radiometric Calibration: FLAME 3 processes each JPEG/TIFF using manufacturer-provided gain/offset constants, extracting per-pixel spectral radiance and inverting Planck’s law for absolute temperature. All steps leverage embedded EXIF metadata for emissivity, atmospheric corrections, and geometric registration, with refined RGB→thermal alignment (<20 px error typical) (Hopkins et al., 2024).

3. Dataset Structure, Organization, and Access

The datasets employ a rigorous, consistent directory and file naming schema:

• FLAME 3 Directory Structure: Separates by angle (“oblique” vs. “nadir”), burn site, and class (fire/nofire), e.g.,

  FLAME3/ ├─ oblique/ │ ├─ fire/ │ └─ nofire/ ├─ nadir/ ├─ preburn_sfm/ ├─ postburn_sfm/ └─ metadata/

• Files: Image pairs named with burn site, timestamp, and modality. Outputs include: radiometric thermal TIFFs (pixel temp in K or °C), thermal JPEGs (inferno colormap), registered RGB–thermal pairs, pre-/post-burn pointclouds, weather logs, and burn site metadata (Hopkins et al., 2024).
• Splits:
  • Public Computer Vision Subset (Sycan Marsh): 738 RGB–TIFF fire pairs, 738 no-fire pairs.
  • Modeling Subset (Hanna Hammock): 800 thermal nadir + 400 RGB frames with GCPs.
  • Full six-burn set available via request for research purposes; subsets hosted on Kaggle.

Earlier releases (FLAME 1/2) are also available under open IEEE Dataport terms (Shamsoshoara et al., 2020). Frame-level and pixel-level annotations (classification and segmentation) are present, with segmentation masks either manually drawn or radiometrically thresholded (FLAME 3, in preparation).

4. Annotation, Baseline Benchmarks, and Applications

Annotation Schemas

• Classification: Frame-wise binary labels (fire/no-fire) determined by manual review and radiometric thresholds (e.g., max $T < 80\,^{\circ}\mathrm{C}$ → “No Fire”, max $T > 200\,^{\circ}\mathrm{C}$ → “Fire”) (Hopkins et al., 2024).
• Segmentation: Pixel-wise binary masks, either hand-annotated (earlier sets) or generated with radiometric thresholds (FLAME 3, under manual check).

Baseline Models and Performance

• Binary Classification
  • FLAME 1/2: Small Xception variant (depth-wise separable convolutions), accuracy: train/val 96.8%/94.3%, test (hold-out UAV): 76.2% (Shamsoshoara et al., 2020).
  • FLAME 3 (Radiometric): RGB+TIFF fusion model yields up to 90.95% accuracy, TIFF-only model up to 91.38% (Hopkins et al., 2024).
• Segmentation
  • Custom U-Net (encoder–decoder, ELU activations, dropout): precision 91.99%, recall 83.88%, F1-score 87.75%, mean IoU 78.17% (Shamsoshoara et al., 2020).
• Modeling, Analysis, Fusion
  • FLAME 3 enables hotspot segmentation, temperature regression, severity mapping, and supports both early/late-fusion multi-modal CNNs and pixelwise regression with Dice, cross-entropy, or MSE loss (Hopkins et al., 2024).

5. Use in Lightweight and Real-Time UAV AI Pipelines

Recent work applies the FLAME datasets to optimize computation for UAV onboard inference. "Streamlining Forest Wildfire Surveillance" (Zhao et al., 2024) restructures FLAME into temporally contiguous 64-frame video clips, applies policy networks (AccSampler) for frame salience, and demonstrates that retaining salient subsets (as few as 20/64 frames) allows heavy-weight video models to achieve >91% accuracy with a >13× reduction in GFLOPs. This confirms the suitability of FLAME for computationally efficient, real-time ML pipelines in disaster response UAV fleets.

6. Limitations, Extensions, and Best Practices

Known Limitations

• Most burns are prescribed and low–moderate intensity. Large, high-intensity wildfires are not yet represented (Hopkins et al., 2024).
• Pixel alignment between RGB and thermal remains imperfect (typical error ~20 px), and missing modalities/measures at some burns.
• Thermal microbolometer arrays used in current sensors exhibit increased noise at extremes of temperature.

Planned or Suggested Extensions

• Inclusion of additional spectral bands (e.g., SWIR, multispectral indices), and active wildland fire collection to broaden coverage.
• Automated per-frame lens distortion correction and stereo rectification.
• Higher cadence nadir sampling (≥1 Hz), pixelwise mask generation via thresholding and manual refinement, addition of bounding boxes and burn-severity polygons.

Recommended Practices

• Always store raw rJPEGs and full EXIF metadata to make future recalibration or correction possible.
• Normalize modalities (RGB and thermal) independently; employ robust ground control plate (GCP) georeferencing (Hopkins et al., 2024).

7. Impact and Research Ecosystem

The FLAME family provides the only open-access, multi-burn, radiometric UAV wildfire datasets with synchronized high-res RGB and per-pixel thermal ground truths. Its adoption enables progress not only in computer vision (detection, segmentation, fusion architectures) but also in fire modeling (rate-of-spread, severity), optimization of edge-compute UAV deployments, and domain adaptation studies. The consistent annotation, open-access subsets, and comprehensive calibration raise the bar for real-world AI applications in wildfire science and disaster response (Shamsoshoara et al., 2020, Hopkins et al., 2024, Zhao et al., 2024).

Researchers can obtain FLAME data for noncommercial research from designated repositories or by contacting the corresponding authors for full-scale datasets. The use of radiometric thermal TIFFs for ground-truth temperature regression is a distinguishing technical advancement, facilitating new research on multimodal fusion, pixelwise estimation, and comprehensive wildfire scene understanding.