PHDataset v1.3

Download all 9,475 patch pairs or choose a city below. Raster data are distributed in 26 independent city ZIPs; each ZIP contains only optical and height GeoTIFF patches. Split lists, normalization parameters and metadata are separate files. The data and benchmark protocol remain v1.3; this is a packaging update.

The Dataset Viewer exposes a searchable patch index and the city/scene tables. It does not decode the seven-band float32 GeoTIFFs as ordinary RGB images. load_dataset loads these metadata tables; download the archives to obtain the rasters.

Dataset summary

PHDataset v1.3 contains 9,475 co-registered image–height patch pairs from 30 PhiSat-2 scenes in 26 cities. It supports monocular building-height estimation and joint height/footprint learning from seven-band optical imagery. Each patch has 256 × 256 pixels on a 4.75 m projected grid.

This release uses a fixed city-disjoint split. All scenes and patches from a city belong to exactly one subset, including patches with local edge overlap. Use the supplied split lists without randomly repartitioning the patches.

Subset	Cities	Patch pairs
Training	16	5,699
Validation	5	2,056
Testing	5	1,720
Total	26	9,475

Validation cities: Beijing, London, Los Angeles, Melbourne, and Rio de Janeiro.

Test cities: Ankara, Las Vegas, Perth, São Paulo, and Tunis.

The remaining 16 cities form the training subset. The supplied splits/v1.3/city_assignments.csv records the complete city allocation. Version v1.3 identifies the dataset and protocol used by the revised TSONet study.

Repository structure

README.md
dataset.json
samples.png
checksums.sha256
data/
  Boston.zip                    # scene directories with TIFF pairs only
  ...                           # one independent archive per city
splits/v1.3/
  train.txt
  val.txt
  test.txt
  city_assignments.csv
normalization/
  normalization_stats.npz
  normalization.json
metadata/
  city_archives.csv
  scene_statistics.csv
  statistics_provenance.json
  patches_train.csv
  patches_val.csv
  patches_test.csv
  FIELDS.md
tools/
  prepare_dataset.py

Every city ZIP has this layout, without a split or version directory inside it:

Boston_3074/
  optical_patches/*.tif
  height_patches/*.tif

Cities with multiple scenes retain separate scene directories inside the same city ZIP. Each split-list line is an optical path relative to the extracted dataset root, not relative to the text file. Replace optical_patches with height_patches to locate its paired label. All scene names and filenames are preserved.

Optical patches contain seven float32 bands; height patches contain one float32 band. GeoTIFF spatial metadata are preserved. Archive SHA-256 values appear in metadata/city_archives.csv; individual TIFF SHA-256 values appear in the patch indices. checksums.sha256 covers every other published repository file.

Download all cities

Install the official CLI once:

python -m pip install -U huggingface_hub hf_xet

Download the release files (the include patterns also avoid unrelated files in the repository):

hf download Yanjiao-WHU/PHDataset --repo-type dataset --local-dir PHDataset_v1.3_download --include "data/*.zip" --include "metadata/*" --include "splits/*" --include "normalization/*" --include "tools/*" --include "dataset.json" --include "README.md" --include "samples.png" --include "checksums.sha256"
python PHDataset_v1.3_download/tools/prepare_dataset.py --output PHDataset_v1.3

The extraction tool verifies archive and raster hashes, extracts all cities to one root, and copies the canonical split lists and normalization files to the root expected by TSONet v1.3. It needs only Python 3.9+ and its standard library. Rerunning it checks and reuses matching existing files and refuses conflicting rasters. The extracted TIFFs occupy approximately 12.70 GB. Allow at least 20 GB of free space for the downloaded archives, extracted data and temporary files.

Download one city

For Boston (262 patch pairs), either use the download link below or run:

hf download Yanjiao-WHU/PHDataset --repo-type dataset --local-dir PHDataset_v1.3_download --include "data/Boston.zip" --include "metadata/*" --include "splits/*" --include "normalization/*" --include "tools/*" --include "dataset.json" --include "README.md" --include "checksums.sha256"
python PHDataset_v1.3_download/tools/prepare_dataset.py --city Boston --output PHDataset_Boston

For several cities, include their archive paths in the download command and repeat --city when extracting, for example --city Boston --city London. City IDs are case-sensitive and match ZIP filenames. City-only extraction does not install complete-benchmark split files; selection.json records the selected cities. Download all cities for the published TSONet training/evaluation protocol.

You may also extract any city ZIP with a standard ZIP utility. Extract all archives to the same directory; do not create an extra city wrapper folder. For manual full preparation, copy dataset.json, splits/v1.3/{train,val,test}.txt, and both normalization/ files to that extracted root.

City downloads

Archive sizes use decimal MB. Total archive size is approximately 5.93 GB. Splitting the release makes selective downloads smaller; it does not substantially reduce the full dataset size.

City	City ID	Split	Scenes	Patch pairs	ZIP size	Download
Adelaide	Adelaide	train	1	305	157.8 MB	ZIP
Alexandria	Alexandria	train	1	237	136.8 MB	ZIP
Ankara	Ankara	test	1	394	301.8 MB	ZIP
Barcelona	Barcelona	train	2	633	481.3 MB	ZIP
Barranquilla	Barranquilla	train	1	124	76.7 MB	ZIP
Beijing	Beijing	val	2	874	411.3 MB	ZIP
Boston	Boston	train	1	262	182.9 MB	ZIP
Cape Town	CapeTown	train	1	341	160.1 MB	ZIP
Constantine	Constantine	train	1	291	211.1 MB	ZIP
Las Vegas	LasVegas	test	1	354	291.4 MB	ZIP
Liuzhou	Liuzhou	train	1	235	181.1 MB	ZIP
London	London	val	1	500	360.0 MB	ZIP
Los Angeles	LosAngeles	val	1	272	157.6 MB	ZIP
Madrid	Madrid	train	1	495	328.8 MB	ZIP
Melbourne	Melbourne	val	1	217	95.4 MB	ZIP
Osaka	Osaka	train	1	500	425.1 MB	ZIP
Ouargla	Ouargla	train	1	252	194.7 MB	ZIP
Perth	Perth	test	1	228	102.8 MB	ZIP
Rio de Janeiro	RiodeJaneiro	val	1	193	110.9 MB	ZIP
Sacramento	Sacramento	train	1	382	205.6 MB	ZIP
Salt Lake City	SaltLakeCity	train	2	635	427.9 MB	ZIP
Santiago	Santiago	train	1	371	190.4 MB	ZIP
São Paulo	SaoPaulo	test	2	506	318.2 MB	ZIP
Tianjin	Tianjin	train	1	490	202.2 MB	ZIP
Tunis	Tunis	test	1	238	115.8 MB	ZIP
Vienna	Vienna	train	1	146	106.8 MB	ZIP

Scene statistics from the manuscript

The following 30-scene inventory reproduces the values reported in the manuscript's dataset table. Download the machine-readable scene statistics and field definitions. Dates are PhiSat-2 acquisition dates in UTC; reference epochs describe the label products.

B is the percentage of jointly valid image-reference pixels with reference height > 2 m. Mean ± SD is the pixel-pooled mean and population standard deviation of those building heights across retained patches. Overlapping locations are included. These are pixel statistics, not building-object statistics. Values are published to two decimal places; no city-level mean is inferred by averaging scene means.

M = Microsoft Global ML Building Footprints and associated heights; G = 3D-GloBFP; E = EUBUCCO v0.1.

City	Scene ID	Acquisition (UTC)	Source	Reference epoch	Split	Pairs	B (%)	Mean ± SD (m)
Adelaide	Adelaide_2959	2025-08-03	M	2014-2024	train	305	10.99	5.36 ± 2.09
Alexandria	Alexandria_3201	2025-08-31	G	2020	train	237	9.37	11.86 ± 6.29
Ankara	Ankara_3145	2025-08-23	G	2020	test	394	6.63	18.62 ± 11.76
Barcelona	Barcelona_2363	2025-05-25	M	2014-2024	train	192	24.64	11.98 ± 4.86
Barcelona	Barcelona_2660	2025-06-24	M	2014-2024	train	441	5.68	7.71 ± 3.70
Barranquilla	Barranquilla_2118	2025-04-25	G	2020	train	124	21.07	11.12 ± 6.81
Beijing	Beijing_2142	2025-04-28	G	2020	val	485	12.56	16.18 ± 10.66
Beijing	Beijing_2560	2025-06-15	G	2020	val	389	12.39	12.79 ± 7.94
Boston	Boston_3074	2025-08-13	M	2014-2024	train	262	11.53	9.58 ± 5.37
Cape Town	CapeTown_2204	2025-05-05	G	2020	train	341	6.34	8.42 ± 6.34
Constantine	Constantine_3708	2025-10-30	G	2020	train	291	5.36	10.99 ± 6.27
Las Vegas	LasVegas_3085	2025-08-14	M	2014-2024	test	354	11.47	5.40 ± 2.42
Liuzhou	Liuzhou_3682	2025-10-28	G	2020	train	235	6.90	23.66 ± 8.77
London	London_1945	2025-04-06	M	2014-2024	val	500	19.71	8.46 ± 3.97
Los Angeles	LosAngeles_3222	2025-09-03	M	2014-2024	val	272	18.63	5.32 ± 2.32
Madrid	Madrid_1543	2024-12-22	M	2014-2024	train	495	16.69	10.96 ± 4.87
Melbourne	Melbourne_2641	2025-06-23	M	2014-2024	val	217	16.07	5.37 ± 1.82
Osaka	Osaka_2015	2025-04-14	G	2020	train	500	20.10	16.68 ± 14.73
Ouargla	Ouargla_3675	2025-10-26	G	2020	train	252	4.87	8.18 ± 3.65
Perth	Perth_2677	2025-06-26	M	2014-2024	test	228	10.81	5.10 ± 1.47
Rio de Janeiro	RiodeJaneiro_2184	2025-05-02	G	2020	val	193	12.11	22.96 ± 16.10
Sacramento	Sacramento_3173	2025-08-25	M	2014-2024	train	382	16.38	5.25 ± 2.16
Salt Lake City	SaltLakeCity_2997	2025-08-06	M	2014-2024	train	425	8.24	5.40 ± 1.83
Salt Lake City	SaltLakeCity_3043	2025-08-10	M	2014-2024	train	210	1.37	5.31 ± 1.69
Santiago	Santiago_2590	2025-06-18	G	2020	train	371	15.15	7.73 ± 4.59
São Paulo	SaoPaulo_1819	2025-03-22	G	2020	test	265	35.66	11.62 ± 7.92
São Paulo	SaoPaulo_1856	2025-03-27	G	2020	test	241	32.11	12.00 ± 8.25
Tianjin	Tianjin_1940	2025-04-06	G	2020	train	490	9.20	11.53 ± 7.17
Tunis	Tunis_3092	2025-08-15	G	2020	test	238	10.10	9.68 ± 4.92
Vienna	Vienna_2133	2025-04-27	E	2013	train	146	15.56	14.16 ± 9.18

Metadata access with Datasets

from datasets import load_dataset
index = load_dataset("Yanjiao-WHU/PHDataset", "patch_index")
boston = index["train"].filter(lambda row: row["city_id"] == "Boston")
scenes = load_dataset("Yanjiao-WHU/PHDataset", "scene_statistics", split="scenes")

These calls return metadata rows and paths within the city ZIPs, not decoded image tensors. The Viewer uses validation for the canonical val.txt subset. For reproducible experiments, use a published release tag or commit hash with revision= (or hf download --revision), and retain the supplied split and file checksums.

Imagery and height references

The retained PhiSat-2 acquisitions span 22 December 2024 to 30 October 2025. Scenes were screened for cloud cover and geometric quality. Seven spectral bands were composited and georeferenced against corresponding landmarks in high-resolution Google Earth imagery, using a common transformation for the bands of each scene. Images and references were placed on an EPSG:3857 grid with 4.75 m pixel spacing.

The imagery contains seven bands in the stored sensor order, beginning with blue, green, and red. TIFF values are retained as supplied in the processed patches; normalization is applied when loading them, not permanently written into the rasters.

Height references are derived from regionally available public products:

• Microsoft Global ML Building Footprints and associated height estimates for North America, Australia, and most European study cities.
• EUBUCCO v0.1 for Vienna.
• 3D-GloBFP for the Asian, South American, and African study cities.

These references differ in source, epoch, and uncertainty. They are not contemporaneous ground surveys of every building. Vector references were rasterized and aligned with the imagery. Height values are in meters.

Valid pixels, normalization, and evaluation

• NoData is −1. A valid image pixel must be valid in all seven bands. Height supervision and evaluation use the intersection of image and height validity; NoData must not be treated as zero-height background.
• Valid zero-height pixels represent background. Continuous reference heights below or equal to 2 m remain valid regression targets.
• normalization/normalization_stats.npz supplies the training-only scalar minimum and maximum across the seven bands: opt_min = 0, opt_max = 4095, and nodata_value = -1. Normalize valid optical values as (value - opt_min) / (opt_max - opt_min); set invalid optical values to zero after masking. Do not fit normalization on validation or test data. No clipping is applied by this normalization.
• The TSONet protocol defines reference buildings by height > 2 m and derives predicted footprints from the final predicted height > 2 m. Building-height errors include missed buildings whose predicted height is ≤2 m. Auxiliary footprint-head scores are a separate diagnostic.
• Model selection uses validation building RMSE. The test subset is reserved for evaluation. The published repeated runs vary the training seed (42–46), while keeping these split lists fixed.

Geographic coverage

Region	Cities
Europe	Barcelona, London, Madrid, Vienna
North America	Boston, Las Vegas, Los Angeles, Sacramento, Salt Lake City
Australia	Adelaide, Melbourne, Perth
Asia	Ankara, Beijing, Liuzhou, Osaka, Tianjin
South America	Barranquilla, Rio de Janeiro, Santiago, São Paulo
Africa	Alexandria, Cape Town, Constantine, Ouargla, Tunis

Intended use and source attribution

PHDataset supports research on image-based building-height mapping, geographically separated evaluation, and joint height/footprint prediction. It should not be treated as authoritative surveyed heights for operational or legal decisions. Refer to the original imagery and reference-product providers for their applicable use and redistribution terms; this dataset card does not replace those terms.

The corresponding TSONet v1.3 implementation is available at SpectorSong/TSONet. Please acknowledge PhiSat-2 and the Microsoft, EUBUCCO, and 3D-GloBFP sources when using the dataset.

Citation

Please cite the associated study and identify PHDataset v1.3 in your experiment description:

@misc{song2026assessmentphisat2imagerymonocular,
  title={The First Assessment of PhiSat-2 Imagery for Monocular Building Height Estimation},
  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}
}