# Hyperspectral and Multispectral Smoke Segmentation Datasets


## Overview

Smoke segmentation is critical for wildfire management and industrial safety applications. Traditional visible-light-based methods face limitations due to insufficient spectral information, particularly struggling with **cloud interference** and **semi-transparent smoke regions**.

![Motivation](figures/figure1_resub.png)
*Figure 1: Motivation for hyperspectral smoke segmentation. The upper part shows challenging smoke scenarios with cloud interference and semi-transparent regions in the visible light band. The lower part plots the spectral distribution of marked points, where yellow shaded regions highlight key discriminatory band ranges.*

To address these challenges, we introduce:
1. **HSSDataset**: The first hyperspectral smoke segmentation dataset with 25 spectral bands
2. **MSSDataset**: A multispectral dataset with RGB-infrared paired images

---

## HSSDataset: Hyperspectral Smoke Segmentation Dataset

### Dataset Overview

HSSDataset is the **first hyperspectral smoke segmentation dataset**, built from an extensive collection of over 18,000 hyperspectral video frames captured across 20 real-world smoke scenarios. From this large-scale raw data collection, we carefully selected and annotated **1,007 high-quality samples** under diverse challenging conditions.

### Hyperspectral Camera System

![Camera Specifications](figures/figure2.png)
*Figure 2: XIMEA MQ022HG-IM-SM5X5-NIR hyperspectral camera specifications and 25-band mosaic filter design.*

We employ a **XIMEA MQ022HG-IM-SM5X5-NIR** hyperspectral camera with the following specifications:

- **Spectral Bands**: 25 bands spanning 600-974nm
- **Wavelengths (nm)**: 600, 616, 632, 647, 664, 680, 696, 712, 728, 744, 760, 776, 792, 808, 824, 840, 856, 872, 888, 894, 910, 926, 942, 958, 974
- **Filter Design**: Specialized 5×5 mosaic filter array deployed on the sensor surface
- **Technology**: Frame-style hyperspectral acquisition based on mosaic coating

**Key Advantages**:
- Simultaneous balance of spectral and spatial resolution
- Rapid acquisition of both spectral and spatial information
- High integration density
- Ability to capture unique "spectral fingerprints" of smoke

### Data Collection

![Challenging Scenarios](figures/figure3.png)
*Figure 3: The challenging scenarios of HSSDataset.*

Data collection was conducted across **20 real-world industrial emission scenarios**, capturing over **18,000 hyperspectral video frames**. The collection targeted diverse smoke conditions, including:

#### Scene-based Challenging Scenarios:
- **High Exposure Environments**: Bright lighting conditions with overexposed regions (214 samples)
- **Low Visibility Conditions**: Poor lighting and atmospheric conditions (118 samples)
- **Complex Backgrounds**: Industrial environments with cluttered backgrounds (411 samples)
- **Cloud Interference**: Scenes with cloud-smoke confusion scenarios (264 samples)

#### Smoke-based Challenging Scenarios:
- **Early-stage Minimal Smoke**: Small smoke plumes in initial emission phases (268 samples)
- **Semi-transparent Regions**: Varying smoke opacity and transparency (258 samples)
- **Complex-shaped Smoke**: Irregular smoke patterns with unclear boundaries (481 samples)

### Annotation Protocol: Many-to-One Annotations

![Annotation Process](figures/figure4.png)
*Figure 4: Many-to-One annotations for hyperspectral smoke segmentation.*

To ensure annotation reliability and capture the inherent uncertainty in smoke boundary delineation, we employ a rigorous **Many-to-One annotations** protocol:

#### Process:
1. **Sampling Strategy**: Systematically sample every 18th frame from the 18,000+ collected frames
2. **Band-Averaged Image Generation**: Generate grayscale images by computing arithmetic mean across all 25 spectral bands
3. **Multiple Annotators**: Each frame receives three independent ground truth masks from three different expert annotators
4. **Ground Truth Definition**: Final masks generated through **majority voting** - each pixel classified as smoke if at least two-thirds of annotators label it as smoke

#### Inter-Annotator Agreement:
- **Unanimous agreement (3/3)**: 52.07% of annotated smoke pixels
- **Majority agreement (2/3)**: 14.14% of annotated smoke pixels
- **Single annotator (1/3)**: 33.79% (excluded from final ground truth)

#### Quality Control:
Special emphasis placed on challenging regions including:
- Early-stage minimal smoke
- Semi-transparent regions
- Blurred boundaries
- Cloud interference areas

### Dataset Structure

```
HSS_VOC/
├── npy_multichannel/          # Hyperspectral data (25 bands)
│   ├── 10_frame_0001.npy
│   ├── 10_frame_0019.npy
│   └── ...
├── Mask/                      # Final ground truth masks 
│   ├── 10_frame_0001.png      # Majority voting results
│   ├── 10_frame_0019.png
│   └── ...
├── Annotation_1/              # Annotator 1's masks 
│   └── *.png
├── Annotation_2/              # Annotator 2's masks 
│   └── *.png
├── Annotation_3/              # Annotator 3's masks 
│   └── *.png
└── split/                     # Dataset split files
    ├── train.txt              # Training set sample list
    ├── val.txt                # Validation set sample list
    └── test.txt               # Test set sample list
```

---

## MSSDataset: Multispectral Smoke Segmentation Dataset

### Dataset Overview

![RGB-IR Pairs](figures/figure5.png)
*Figure 5: Visible and infrared frame pairs in MSSDataset.*

To validate the generalizability of our method beyond hyperspectral data, we constructed a **multispectral smoke segmentation dataset (MSSDataset)** derived from the [FLAME2 dataset](https://ieee-dataport.org/open-access/flame-2-fire-detection-and-modeling-aerial-multi-spectral-image-dataset).

### Dataset Specifications

- **Source**: [FLAME2 dataset](https://ieee-dataport.org/open-access/flame-2-fire-detection-and-modeling-aerial-multi-spectral-image-dataset)
- **Spectral Channels**: 4 channels (RGB + Infrared)
- **Annotated Samples**: 200 carefully selected samples
- **Scenarios**: Wildland fire smoke under various environmental conditions
- **Format**: RGB-IR paired multispectral cubes

### Key Features

- Enhanced infrared visualization for better smoke detection
- Diverse wildfire scenarios
- Distinct smoke features suitable for annotation
- Complementary to HSSDataset for cross-modality validation

### Dataset Structure

```
FLAME2_VOC/
├── npy_multichannel/          # Multispectral data (4 channels: RGB+IR)
│   └── *.npy
├── Mask/                      # Ground truth masks
│   └── *.png
└── split/                     # Dataset split files
    ├── train.txt              # Training set sample list 
    ├── val.txt                # Validation set sample list 
    └── test.txt               # Test set sample list
```


---

## Data Format

### Hyperspectral Data (HSSDataset)

Each `.npy` file contains a hyperspectral image cube:
- **Shape**: `(H, W, 25)` where H and W are spatial dimensions
- **Data Type**: `float32` or `uint16`
- **Channels**: 25 spectral bands (600-974nm)
- **Organization**: Channels ordered by wavelength

Example loading code:
```python
import numpy as np

# Load hyperspectral data
hsi_data = np.load('10_frame_0001.npy')  # Shape: (H, W, 25)

# Access specific spectral band
band_600nm = hsi_data[:, :, 0]  # First band (600nm)
band_974nm = hsi_data[:, :, 24]  # Last band (974nm)

# Generate band-averaged image (for visualization)
band_avg = np.mean(hsi_data, axis=2)
```

### Multispectral Data (MSSDataset)

Each `.npy` file contains a multispectral image:
- **Shape**: `(H, W, 4)` where H and W are spatial dimensions
- **Data Type**: `float32`
- **Channels**: 4 channels (RGB + IR)
- **Channel Order**: `[R, G, B, IR]`

Example loading code:
```python
import numpy as np

# Load multispectral data
msi_data = np.load('sample_001.npy')  # Shape: (H, W, 4)

# Access RGB channels
rgb = msi_data[:, :, :3]

# Access infrared channel
ir = msi_data[:, :, 3]
```

---

## 📝 Citation

If you use these datasets in your research, please cite:

```bibtex
@article{yao2026hyperspectral,
  title={Hyperspectral Smoke Segmentation via Mixture of Prototypes},
  author={Yao, Lujian and Zhao, Haitao and Kong, Xianghai and Xu, Yuhan}
  year={2026},
  journal={arXiv preprint arXiv:2602.10858},
}
```

**Dataset on Hugging Face:** [https://huggingface.co/datasets/LujianYao/HSSDataset](https://huggingface.co/datasets/LujianYao/HSSDataset)


---

## License

[![License: CC BY-NC-SA 4.0](https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc-sa/4.0/)

This dataset is released under the **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)**.For more details, see the full license at: https://creativecommons.org/licenses/by-nc-sa/4.0/

---

## Acknowledgments

- [FLAME2 dataset](https://ieee-dataport.org/open-access/flame-2-fire-detection-and-modeling-aerial-multi-spectral-image-dataset) authors for providing the source data for MSSDataset