MultiFire20K: A semi-supervised enhanced large-scale UAV-based dataset for fire monitoring

MultiFire20K

MultiFire20K is a UAV-based dataset for fire, classification, segmentation and multi-task learning. It contains 20,500 images extracted from 67 videos, covering diverse urban (build-up) and rural (natural landscape) scenes worldwide.

Dataset Stats
- Images: 20,500
- Videos: 67
- Categories: Fire / Normal
- Environments: Urban / Rural

│ Category │ Build-Up │ Natural Landscape  │  Total │
│ Fire          │    5,075   │              5,085           │ 10,160 │
│ Normal    │    5,110   │              5,230           │ 10,340 │
│ Total        │   10,185   │             10,315         │ 20,500 │

Annotations
- Manual masks (~10%) created with Fiji and LabelKit.
- Pseudo-labels (~90%) via a semi-supervised ensemble (for fire images).
- Normal images: RGB only (no masks).

│ Fire Subset            │ Human Ann. │ Pseudo-Labels  │   Total │
│ Build-Up (Urban)   │        510       │          4,565        │   5,075 │
│ Rural (Landscape)  │        515       │          4,570        │   5,085 │
│ TOTAL                    │      1,025       │          9,135        │  10,160 │

File Naming & Traceability
Each image encodes its source video and frame number:  
`FramesV1_1000.jpg` → extracted from Video 1, frame 1000.

Image & Mask Formats
- Images: `.jpg` (various resolutions; mean ≈ 1147×636, range 426–1920 × 210–1080)
- Masks: `.tif` (single-channel, pixel-aligned with the image, used for *fire* images only)

Folder Structure
MultiFire20K/
├─ Annotations/
│  ├─ Fire_Rural_Annotate/        # fire rural: image.jpg + mask.tif
│  ├─ Fire_Rural_Annotate_RGB/    # fire rural RGB only
│  ├─ Fire_Urban_Annotate/        # fire urban: image.jpg + mask.tif
│  ├─ Fire_Urban_Annotate_RGB/    # fire urban RGB only
├─ Train/                         # FR, FU, NR, NU subfolders
├─ Val/                  # FR, FU, NR, NU subfolders
├─ Test/               # FR, FU, NR, NU subfolders
└─ data_structure.csv            # metadata (label_type, fire_type, category, split)
```

Split codes:  
> FR = Fire Rural, FU = Fire Urban, NR = Normal Rural, NU = Normal Urban.  
> Rural is also known as Natural Landscape; Urban as Build-Up.

Note: In FR and FU (fire) folders you’ll find image + mask pairs (.jpg + .tif). In NR and NU (normal) folders there are images only (.jpg, no masks). 
Use data_structure.csv to check whether a fire mask is manual or pseudo.

Metadata (`data_structure.csv`)
Columns:
- `image_name` — filename (e.g., `FramesV11_24900.jpg`)
- `label_type` — `manual` (human) or `pseudo` (generated)
- `fire_type` — `fire` or `normal`
- `category` — `rural` (natural landscape) or `urban` (build-up)
- `split` — `train`, `val`, `test`

Use this table to:
- filter manual vs pseudo masks,
- select fire vs normal,
- choose rural vs urban,
- load the official split.

Recommended Usage

Classification
- Labels: `fire_type` (binary) and optionally `category` (urban/rural).
- Use FR/FU/NR/NU folders or `data_structure.csv` for labels and splits.

Segmentation
- Pair each `image.jpg` with `mask.tif` of the same basename (fire images only).
- Prefer manual** masks for evaluation; pseudo-labels can be used for training at scale.

Citation
If you use MultiFire20K, please cite:

@article{shianios2025multifire20k,
  title={MultiFire20K: A semi-supervised enhanced large-scale UAV-based benchmark for advancing multi-task learning in fire monitoring},
  author={Shianios, Demetris and Kolios, Panayiotis and Kyrkou, Christos},
  journal={Computer Vision and Image Understanding},
  volume={254},
  pages={104318},
  year={2025},
  publisher={Elsevier}
}