KwasiKodie/Puerto-Rico-Building-Damage-Segmentation

Puerto Rico Building Damage Segmentation

Dataset Description

Puerto Rico Building Damage Segmentation is a geospatial computer-vision dataset containing high-resolution satellite image tiles with polygon annotations for identifying building type and apparent storm-damage state.

The dataset was developed from work undertaken in connection with the 2024 EY Open Science Data Challenge: Coastal Resilience.

The underlying satellite imagery was provided through the Maxar Open Data Program and originates from GEO-1 imagery covering areas of Puerto Rico affected by Hurricane Maria.

The publication release contains 4,200 labelled satellite image tiles and 32,981 annotated building instances.

This is a cleaned publication copy of the original research dataset. The original research files were preserved separately and were not modified during publication preparation.

Dataset Details

Task

The primary task is polygon-based building segmentation combined with building-type and apparent damage-state classification.

A single image can contain multiple buildings and multiple classes. Therefore, this is not a conventional single-label image-classification dataset.

Classes

ID	Class	Description
0	damaged_residential_building	Residential building annotated as damaged
1	undamaged_residential_building	Residential building annotated as undamaged
2	undamaged_commercial_building	Commercial building annotated as undamaged
3	damaged_commercial_building	Commercial building annotated as damaged

Dataset Statistics

Split	Images	Building instances
Train	3,145	24,772
Validation	844	6,641
Test	211	1,568
Total	4,200	32,981

Class Distribution

ID	Class	Instances	Share
0	damaged_residential_building	23,276	70.57%
1	undamaged_residential_building	7,435	22.54%
2	undamaged_commercial_building	1,829	5.55%
3	damaged_commercial_building	441	1.34%
	Total	32,981	100%

The dataset is strongly class-imbalanced. In particular, damaged_commercial_building represents only about 1.34% of annotated instances.

Dataset Structure

The public release is organized as follows:

YOLODataset_HF/
├── README.md
├── LICENSE_DATA.md
├── data.yaml
├── cleaning_report.json
├── removed_cross_split_duplicates.csv
├── final_audit.json
├── images/
│   ├── train/
│   ├── val/
│   └── test/
└── labels/
    ├── train/
    ├── val/
    └── test/

Each labelled image has a corresponding annotation file with the same filename stem.

For example:

images/train/example.png
labels/train/example.txt

The local publication workspace also contains a submission/ directory with 12 challenge evaluation images. These images are intentionally excluded from the planned public Hugging Face release pending separate confirmation of their redistribution terms.

Annotation Format

Annotations use YOLO polygon-segmentation format.

Each non-empty line represents one building instance:

<class_id> <x1> <y1> <x2> <y2> ... <xn> <yn>

The first value is the integer class identifier.

The remaining values are alternating x and y coordinates describing the vertices of the building polygon.

All coordinates are normalized to the interval [0,1] relative to the image width and height.

A polygon contains at least three coordinate pairs. Because an image can contain multiple buildings, a label file can contain multiple annotation lines and multiple classes.

Class mapping

0 damaged_residential_building
1 undamaged_residential_building
2 undamaged_commercial_building
3 damaged_commercial_building

The same mapping is provided in data.yaml.

Image Characteristics

All 4,200 labelled images in the publication dataset are PNG files.

Most tiles have dimensions of 512 × 512 pixels.

The final audit found the following dimensions:

Split	512 × 512	371 × 512	Total
Train	3,121	24	3,145
Validation	833	11	844
Test	208	3	211
Total	4,162	38	4,200

The 38 images measuring 371 × 512 pixels are retained edge tiles from the source-image tiling process. They were deliberately preserved rather than resized, padded or discarded.

No image resizing or conversion was performed as part of publication cleaning.

Source Data and Provenance

The dataset was developed from work undertaken in connection with the 2024 EY Open Science Data Challenge: Coastal Resilience.

The challenge provided high-resolution satellite imagery for analysing coastal infrastructure and storm damage in Puerto Rico.

According to the challenge participant materials, the underlying imagery was supplied through the Maxar Open Data Program and included GEO-1 imagery.

The research task involved identifying individual structures and distinguishing them according to building type and apparent damage state.

This repository does not represent the original Maxar imagery products in their original form. It contains image tiles and polygon annotations prepared for machine-learning research.

Disaster context

The imagery used for the research relates to areas of Puerto Rico affected by Hurricane Maria.

The challenge materials provided pre-event and post-event imagery for the broader research task. The labelled image tiles in this publication were prepared for building-damage analysis using that challenge imagery.

Users should therefore interpret the labels as remote-sensing annotations within this specific disaster-analysis context rather than as independent structural-engineering assessments.

Data Preparation

The source satellite imagery was divided into smaller image tiles for computer-vision experimentation.

The principal tile size was 512 × 512 pixels. Edge regions generated a small number of 371 × 512 pixel tiles, which remain in the dataset.

Individual visible buildings were represented using polygon boundaries.

Each polygon was assigned one of four class identifiers representing the combination of:

• residential or commercial building type; and
• damaged or undamaged apparent damage state.

The resulting labels were stored using normalized YOLO polygon coordinates.

Original Research Dataset

Before publication cleaning, the labelled research dataset contained:

Split	Images
Train	3,171
Validation	845
Test	211
Total	4,227

The original dataset contained 33,219 annotated building instances.

Its original class distribution was:

Class	Instances
damaged_residential_building	23,450
undamaged_residential_building	7,489
undamaged_commercial_building	1,837
damaged_commercial_building	443
Total	33,219

The original files were preserved separately. Publication cleaning was performed on a copy so that the source research dataset remained unchanged.

Publication Cleaning

A systematic integrity audit was performed before preparing the dataset for public distribution.

The audit examined:

• image and label counts;
• filename correspondence;
• empty labels;
• orphan labels;
• polygon structure;
• class identifiers;
• coordinate validity;
• image dimensions and formats;
• exact image duplication using SHA-256; and
• duplication across train, validation and test splits.

Cross-split duplicate detection

The original audit identified 27 exact image hash groups that crossed dataset split boundaries.

The duplicate pattern was:

• 20 groups crossing train and validation;
• 6 groups crossing train and test; and
• 1 group crossing validation and test.

Exact copies appearing in both training and evaluation splits can produce information leakage and artificially influence model evaluation.

For the publication copy, a deterministic retention priority was used:

test > validation > train

This preserves evaluation samples in preference to copies appearing in training data.

The procedure removed:

Split	Removed image/label pairs
Train	26
Validation	1
Test	0
Total	27

The resulting publication split sizes are:

Split	Original	Removed	Published
Train	3,171	26	3,145
Validation	845	1	844
Test	211	0	211
Total	4,227	27	4,200

The corresponding annotation count changed from 33,219 building instances in the original research copy to 32,981 instances in the publication copy.

Coordinate quality control

The original annotation audit identified 186 coordinate values marginally outside the normalized interval because they were infinitesimally below zero.

All 186 values were negative floating-point artifacts. No values greater than 1 were detected.

The most negative offending value was approximately:

-1.1102230246251565e-16

Such a value is effectively zero at the precision relevant to normalized image coordinates.

Of the 186 affected coordinate values:

• 184 occurred in annotations retained in the publication dataset and were clamped to exactly 0;
• 2 occurred in annotation files associated with duplicate samples removed during cross-split cleaning.

After cleaning, the independent audit found zero coordinates outside the interval [0,1].

Cache exclusion

YOLO-generated cache files were not included in the publication copy.

Files such as:

train.cache
val.cache

are runtime artifacts rather than source annotations and are unnecessary for reproducing the dataset.

Preservation of source imagery

Publication cleaning did not resize, crop, recompress or otherwise alter the retained image files.

The 371 × 512 edge tiles were intentionally retained.

The original research dataset remained unchanged throughout this process.

Quality Assurance

After publication cleaning, the complete dataset was independently re-audited.

The final audit examined every retained image and annotation file and confirmed:

Integrity check	Result
Images	4,200
Label files	4,200
Building instances	32,981
Missing labels	0
Orphan labels	0
Empty labels	0
Invalid annotations or unreadable images	0
Cross-split exact duplicate groups	0

All four expected class identifiers (0 through 3) are represented in the final dataset.

The audit also verified that polygon coordinates fall within [0,1] and that annotation records contain valid polygon coordinate structures.

For transparency, the publication directory includes:

• cleaning_report.json — records publication-cleaning operations;
• removed_cross_split_duplicates.csv — identifies duplicate samples excluded from the publication copy; and
• final_audit.json — records the post-cleaning integrity audit.

These files are included to make the publication transformation inspectable rather than presenting the cleaned release as identical to the original research dataset.

Remaining Duplicate Images

The final audit identified 49 exact duplicate hash groups remaining within individual dataset splits.

No exact duplicate group crosses the train, validation and test boundaries.

The remaining within-split duplicates were deliberately retained because the publication-cleaning objective was to eliminate exact cross-split leakage while minimizing unnecessary changes to the research corpus.

Users who require every image in a split to be byte-wise unique may perform an additional within-split deduplication step.

Any such modification should be documented because it will change image and annotation counts relative to this published release.

Intended Uses

The dataset is intended primarily for research and educational use in computer vision, remote sensing and disaster-related geospatial analysis.

Potential uses include:

• building instance segmentation;
• post-disaster building assessment research;
• remote-sensing computer vision;
• geospatial machine learning;
• satellite-image analysis;
• disaster-response research;
• building-type recognition;
• apparent building-damage classification;
• class-imbalanced segmentation research;
• evaluation of segmentation architectures;
• transfer-learning experiments; and
• reproducibility studies.

The dataset may also support research into methods that jointly model building type and apparent damage state.

Out-of-Scope Uses

The dataset should not be treated as a substitute for professional on-site inspection or authoritative disaster assessment.

It should not be used as the sole basis for:

• determining whether a real building is structurally safe;
• deciding whether a structure can safely be occupied;
• emergency evacuation decisions;
• insurance or compensation determinations;
• legal determinations of property damage;
• property valuation;
• individual eligibility decisions; or
• other safety-critical decisions requiring professional assessment.

A visual damage label derived from overhead satellite imagery is not equivalent to a structural engineering assessment.

Limitations

Geographic scope

The dataset represents a specific geographic and disaster context in Puerto Rico.

Buildings, construction practices, roofing materials, settlement patterns, vegetation and surrounding land cover may differ substantially from those found elsewhere.

Model performance on this dataset should therefore not be interpreted as evidence of equivalent performance in other countries or regions.

Disaster-specific context

The underlying imagery is associated with the Hurricane Maria disaster context.

Damage characteristics caused by hurricanes may differ from those produced by earthquakes, floods, fires, conflict or other hazards.

Models trained on this dataset may therefore experience substantial domain shift when applied to different disaster types.

Class imbalance

The final publication dataset contains:

• 23,276 damaged residential instances (70.57%);
• 7,435 undamaged residential instances (22.54%);
• 1,829 undamaged commercial instances (5.55%); and
• 441 damaged commercial instances (1.34%).

The strong imbalance means aggregate metrics can obscure poor performance on minority classes.

Researchers should consider per-class precision, recall, F1 score, intersection-over-union and other class-sensitive measures when evaluating models.

Spatial dependence

Removal of exact cross-split duplicates does not guarantee complete geographic independence between dataset splits.

Individual tiles may originate from nearby areas within larger satellite scenes. Consequently, different splits may contain geographically close locations with similar:

• building morphology;
• roofing materials;
• vegetation;
• road patterns;
• land cover;
• illumination; and
• sensor/acquisition characteristics.

The absence of exact duplicate images should therefore not be interpreted as proof that all spatial autocorrelation or spatial leakage has been eliminated.

For experiments requiring strict geographic independence, researchers should construct geographically separated splits when sufficient spatial provenance is available.

Annotation uncertainty

Damage state is inferred from remotely sensed visual evidence.

Satellite imagery may not reveal:

• internal structural damage;
• damage obscured by roofs or vegetation;
• foundation damage;
• damage not visible at the available spatial resolution; or
• the engineering condition of a structure.

Consequently, apparent visual damage labels may differ from ground-based engineering assessments.

Sensor and acquisition dependence

Models may learn characteristics associated with the particular imagery source, resolution, acquisition conditions and preprocessing workflow.

Performance may change when models are applied to imagery from other satellites, resolutions, dates, seasons, atmospheric conditions or image processing pipelines.

Edge-tile dimensions

Thirty-eight images have dimensions of 371 × 512 pixels rather than 512 × 512 pixels.

Training pipelines that assume fixed image dimensions should explicitly handle these images through suitable resizing, padding, batching or other preprocessing.

Bias and Representation

This dataset is not intended to represent the full distribution of buildings in Puerto Rico or globally.

The dataset contains a strong overrepresentation of damaged residential buildings and relatively few damaged commercial examples.

Models trained without accounting for this imbalance may preferentially learn the dominant classes.

The geographic concentration of the imagery may also cause models to learn location-specific visual characteristics that do not generalize to other environments.

Researchers should therefore report class-specific performance and evaluate external generalization before drawing broader conclusions.

Ethical and Responsible Use Considerations

Automated damage assessment can support rapid analysis of large affected areas, but model predictions can contain false positives and false negatives.

Errors may have significant consequences if predictions are interpreted as authoritative assessments of individual properties.

Human review and appropriate domain expertise should therefore accompany real-world applications, particularly where decisions could affect safety, access to services, financial outcomes or disaster assistance.

The dataset should be used in a manner consistent with the applicable terms governing the underlying satellite imagery.

Licensing and Attribution

Underlying imagery

The underlying satellite imagery originates from the Maxar Open Data Program and was supplied in connection with the 2024 EY Open Science Data Challenge: Coastal Resilience.

The Maxar-derived imagery distributed in this dataset is subject to the applicable Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) terms.

Users should review the applicable source-data terms before redistribution or downstream use, particularly where commercial use is contemplated.

CC BY-NC 4.0 requires, among other conditions, appropriate attribution, identification of modifications where applicable, and compliance with the non-commercial restriction.

Modifications and derived material

This repository contains derived material rather than the original satellite products in their original organization.

Transformations represented in this release include:

• tiling of source imagery for computer-vision workflows;
• polygon annotation of visible building instances;
• assignment of building-type and apparent damage-state classes;
• organization into train, validation and test splits;
• exclusion of exact cross-split duplicate samples;
• normalization of negligible floating-point coordinate artifacts; and
• publication-specific documentation and quality assurance.

The repository includes LICENSE_DATA.md with additional provenance and licensing information.

Suggested attribution

When using or redistributing material from this dataset, users should retain appropriate attribution to the source imagery.

A suggested attribution is:

Contains imagery derived from the Maxar Open Data Program and prepared for building-damage research associated with the 2024 EY Open Science Data Challenge: Coastal Resilience. Imagery was tiled and annotated for machine-learning research. Original research and dataset preparation by Kwasi Gyamfi Kodie, Eric Akuamoah and Nana Kwabena Kodie. Publication cleaning, quality assurance and Hugging Face release preparation by Kwasi Gyamfi Kodie.

Users remain responsible for ensuring that their particular use complies with the applicable terms governing the underlying imagery.

Relationship to the EY Open Science Data Challenge

The dataset was developed from research undertaken in connection with the 2024 EY Open Science Data Challenge: Coastal Resilience.

EY provided the challenge framework and participant materials through which the research task was undertaken.

This Hugging Face repository is an independently prepared research dataset publication.

It should not be interpreted as:

• an official EY dataset;
• an EY-maintained repository;
• an EY certification of the annotations;
• an EY endorsement of models trained using this dataset; or
• an EY endorsement of this Hugging Face publication.

References to EY are included to accurately document the research context and provenance.

Relationship to Maxar

Maxar Technologies / Maxar Open Data Program is the source of the underlying satellite imagery.

The tiling, annotations, dataset organization, publication cleaning and documentation contained in this repository were prepared as part of the research and publication workflow.

These derived elements should not be interpreted as modifications, annotations or quality assessments produced or endorsed by Maxar.

Users of the imagery should preserve the attribution required by the applicable source-data license.

Reproducibility

The repository includes data.yaml, which defines the class mapping used by the YOLO segmentation dataset.

Publication quality-control records are provided in:

cleaning_report.json
removed_cross_split_duplicates.csv
final_audit.json

Together, these files document:

• the publication-cleaning policy;
• the exact cross-split duplicates removed;
• coordinate normalization performed during cleaning; and
• the final post-cleaning integrity audit.

Researchers should record the exact Hugging Face repository revision or commit used in experiments because future releases may contain documentation corrections or additional quality-control improvements.

Dataset Version

This is the initial cleaned release of the research dataset.

Publication statistics for this release are:

Labelled images:       4,200
Building instances:   32,981
Training images:       3,145
Validation images:       844
Test images:             211
Classes:                   4

The original research corpus contained 4,227 labelled images before the 27 cross-split duplicate image/label pairs were excluded from this publication copy.

Citation

If this dataset contributes to academic research, please cite the Hugging Face dataset repository and acknowledge the original Maxar imagery source.

Until a persistent DOI or associated dataset publication is assigned, the repository can be cited using the dataset title, authors, platform and repository URL.

Suggested citation

Kodie, Kwasi Gyamfi; Akuamoah, Eric; Kodie, Nana Kwabena. Puerto Rico Building Damage Segmentation.
Hugging Face Datasets, 2026. https://huggingface.co/datasets/KwasiKodie/Puerto-Rico-Building-Damage-Segmentation

BibTeX template

@dataset{kodie2026puertorico,
  author    = {Kodie, Kwasi Gyamfi and Akuamoah, Eric and Kodie, Nana Kwabena},
  title     = {Puerto Rico Building Damage Segmentation},
  year      = {2026},
  publisher = {Hugging Face},
  note      = {Derived from Maxar Open Data Program imagery and research
               conducted in connection with the 2024 EY Open Science
               Data Challenge: Coastal Resilience},
  url       = {https://huggingface.co/datasets/KwasiKodie/Puerto-Rico-Building-Damage-Segmentation}
}

The citation information above uses the permanent Hugging Face dataset repository address. If a DOI is assigned in the future, the citation can be updated accordingly.

Acknowledgements

Maxar Technologies / Maxar Open Data Program

The underlying high-resolution satellite imagery used in this research originates from the Maxar Open Data Program.

EY Open Science Data Challenge

The 2024 EY Open Science Data Challenge: Coastal Resilience provided the research context and participant materials associated with the original building-damage task.

Research and dataset authorship

The original research and dataset were prepared by Kwasi Gyamfi Kodie, Eric Akuamoah and Nana Kwabena Kodie. Their work included the research workflow, annotation organization and machine-learning experimentation represented by the original research dataset.

For this Hugging Face release, Kwasi Gyamfi Kodie subsequently carried out the publication-specific cleaning, quality assurance, documentation and repository preparation described in this dataset card.

References

The following sources provide important context for the dataset:

1. EY Open Science Data Challenge — 2024 Participant Guidance. Participant documentation describing the Coastal Resilience challenge, Puerto Rico imagery, GEO-1 source data and challenge workflow.

2. EY Open Science Data Challenge — 2024 Terms and Conditions. Terms governing participation in the 2024 challenge.

3. Maxar Open Data Program. Source program for the underlying satellite imagery.

4. Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). License terms applicable to the Maxar Open Data material identified for this release.

Contact and Corrections

Questions about the dataset preparation, annotation organization or publication quality-control process can be raised through the Hugging Face dataset repository after publication.

If annotation errors, duplicate samples or documentation inaccuracies are identified, they should be reported with the relevant image filename and, where applicable, annotation line or class.

Corrections should be released transparently so that changes between dataset versions remain traceable.