PHDataset

Dataset Summary

PHDataset is a global remote sensing dataset for monocular building height estimation from PhiSat-2 optical imagery. It was constructed by combining globally distributed PhiSat-2 scenes with open-source building height labels from multiple sources, followed by manual georeferencing, rasterization, masking, and patch generation.

The dataset is intended to support research on:

• building height estimation from optical satellite imagery,
• ordinal and regression-based dense prediction for urban remote sensing,
• cross-region generalization in global urban structure analysis.

PHDataset contains 9,475 co-registered image-label patch pairs. Each sample is cropped to 256 × 256 pixels. The full dataset is randomly split into training, validation, and testing subsets with a ratio of 7:2:1.

---

Dataset Description

What is this dataset?

PHDataset is a machine learning dataset designed for building height estimation from multispectral optical satellite imagery. The imagery comes from PhiSat-2, while the height labels are compiled from several publicly available building-height data sources. The dataset was created to evaluate the suitability of PhiSat-2 imagery for dense urban height estimation at the global scale.

What problem does it address?

Despite the growing interest in monocular building height estimation, publicly available global-scale datasets that combine recent optical imagery with building height labels remain limited. PHDataset addresses this gap by providing:

• global coverage across multiple continents,
• multispectral imagery with seven bands,
• co-registered height labels suitable for patch-based learning.

Modality and spatial characteristics

• Sensor: PhiSat-2
• Imagery type: multispectral optical imagery
• Number of bands: 7
• Spatial resolution: 4.75 m
• Patch size: 256 × 256 pixels
• Data type: float32

---

Source Data

PhiSat-2 imagery

PhiSat-2 images acquired from launch to 22 October 2025 were collected for multiple urban areas worldwide. After screening for low cloud cover and limited geometric distortion, 30 scenes from 26 cities were retained.

The seven PhiSat-2 bands were composited and reprojected to WGS 1984 Web Mercator (EPSG:3857). Because noticeable spatial offsets were present in the original products, the imagery was manually georeferenced using high-resolution Google Earth imagery as reference. The final imagery was resampled to 4.75 m spatial resolution.

Building height labels

Open building height labels were compiled from multiple sources:

1. Microsoft GlobalMLBuildingFootprints for regions in North America, Australia, and most of Europe;
2. EUBUCCO v0.1 for the Vienna region;
3. 3D-GloBFP for regions in Asia, South America, and Africa.

All vector labels were projected to EPSG:3857, rasterized at 4.75 m, manually aligned to the processed PhiSat-2 imagery, and resampled again to ensure spatial consistency.

---

File Structure

Please adapt this section if your released ZIP file uses slightly different folder names. A typical organization is:

PHDataset/
├── train/
│   ├── image/
│   └── label/
├── val/
│   ├── image/
│   └── label/
├── test/
│   ├── image/
│   └── label/
└── splits/
    ├── train.txt
    ├── val.txt
    └── test.txt

Each sample typically contains:

• a 7-band PhiSat-2 image patch
• a building height label patch
• optional auxiliary files depending on the released package

---

Geographic Coverage

PHDataset covers urban regions from six continents.

Sampled cities by region

Region	Cities
Europe (EU)	Barcelona, London, Madrid, Vienna
North America (NA)	Boston, LasVegas, LosAngeles, Sacramento, SaltLakeCity
Australia (AU)	Adelaide, Melbourne, Perth
Asia (AS)	Ankara, Beijing, Liuzhou, Osaka, Tianjin
South America (SA)	Barranquilla, RiodeJaneiro, Santiago, SaoPaulo
Africa (AF)	Alexandria, CapeTown, Constantine, Ouargla, Tunis

Sampling map

Geographic distribution of sampled cities and PhiSat-2 scenes in PHDataset.

---

Intended Uses

PHDataset is intended for research use in areas such as:

• monocular building height estimation,
• dense urban structure mapping,
• multi-task learning with footprint segmentation and height prediction,
• remote sensing model benchmarking.

Users can also see TSONet as a using example: https://github.com/SpectorSong/TSONet

It is not intended to be treated as authoritative building-height ground truth for operational or legal applications.

---

Citation

If you use PHDataset in your research, please cite the corresponding paper, which is currently under review and available in arXiv:

@misc{song2026monocularbuildingheightestimation,
      title={Monocular Building Height Estimation from PhiSat-2 Imagery: Dataset and Method}, 
      author={Yanjiao Song and Bowen Cai and Timo Balz and Zhenfeng Shao and Neema Simon Sumari and James Magidi and Walter Musakwa},
      year={2026},
      eprint={2603.29245},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.29245}, 
}

---

Data Sources Acknowledgement

PHDataset is built upon PhiSat-2 imagery and open building height labels from:

• Microsoft GlobalMLBuildingFootprints
• EUBUCCO v0.1
• 3D-GloBFP

Please also cite the original data sources when appropriate.

---