---

pretty_name: "Pothole and Road Damage Observations 2023-2025"
license: cc-by-sa-4.0
size_categories: ["10K<n<100K"]
tags: ["geospatial", "open-data", "tabular", "computer-vision", "pothole", "potholes", "road-damage", "road-infrastructure", "road-maintenance", "street-level-imagery", "visual-search", "semantic-search", "image-geolocation", "infrastructure", "urban", "mapping", "parquet", "geojson", "csv", "mapillary", "inaturalist", "year-2023", "year-2024", "year-2025"]
---

# Pothole & Road Damage Dataset — 2023–2025

A geospatial dataset of **10,811 pothole and road-damage observations** collected from geolocated street-level imagery and physical-world observations from 2023, 2024, and 2025.

The dataset was created and distributed by **[EmbedEarth](https://embed.earth)**, programmable geographic infrastructure for searching, retrieving, and computing across the physical world.

Observations were discovered using the semantic query **“Pothole”** and selected using available visual context, source text, metadata, temporal information, and geographic context.

This dataset can be used for **pothole detection, road damage analysis, computer vision, infrastructure monitoring, road maintenance research, geospatial AI, urban planning, and visual search**.

## Search millions more geographic features with EmbedEarth

This Hugging Face dataset represents one downloadable query from the broader **[EmbedEarth](https://embed.earth)** geographic index.

EmbedEarth gives developers and AI systems access to **millions of geographic features and large collections of geolocated images and videos** covering roads, infrastructure, buildings, places, community issues, nature, and other features of the physical world.

Instead of relying only on static downloadable datasets, you can search and work with geographic data programmatically.

### Build with EmbedEarth

* **[Geographic Feature List](https://www.embed.earth/catalog/features)** — browse geographic features available through EmbedEarth
* **[EmbedEarth](https://embed.earth)** — programmable infrastructure for Earth
* **[Developer Documentation](https://www.embed.earth/docs)** — Search, Compute, Routing, Geocoding, Maps, Data, SDK, CLI, and MCP
* **[Geographic Search SDK](https://www.embed.earth/docs/sdk/search)** — query geographic features and regions programmatically
* **API** — add geographic search and spatial data to applications and data pipelines
* **SDK** — build geographic capabilities directly into JavaScript and TypeScript applications
* **CLI** — search and work with physical-world data from the terminal
* **MCP** — give AI agents access to geographic search and spatial tools

## Search the physical world

The same geographic infrastructure used to build this pothole dataset can be used to find many other physical-world conditions and features.

Example searches include:

```text
potholes in New York
potholes in Los Angeles
road damage in California
damaged pavement in Toronto
cracked roads in Chicago
road construction in Manhattan
damaged sidewalks in Brooklyn
traffic accidents in Texas
construction permits in Miami
fallen trees in Florida
crosswalks in San Francisco
utility poles in Houston
```

See the **[EmbedEarth Search SDK documentation](https://www.embed.earth/docs/sdk/search)** to query published geographic features by area, coordinates, bounding box, attributes, and other spatial constraints.

## Dataset overview

This release contains **10,811 geolocated pothole-related observations** covering:

* 2023
* 2024
* 2025

Each record was selected because available imagery, metadata, text, or geographic context matched the pothole query.

Where available, observations include:

* latitude and longitude
* spatial geometry
* observation timestamps
* street-level imagery or media URLs
* source attribution
* source URLs
* source-specific metadata
* temporal context

The dataset is designed for **geographic discovery and analysis** rather than as a complete government or municipal inventory of every pothole.

## What can this pothole dataset be used for?

### Pothole detection

Use the observations as part of research or evaluation workflows for identifying potholes and damaged pavement from street-level imagery.

### Road damage detection

Explore visual and geographic indicators of road deterioration, pavement defects, surface damage, and transportation infrastructure conditions.

### Computer vision

Use pothole observations in computer vision and multimodal research involving street imagery, physical-world understanding, visual retrieval, or image geolocation.

### Road maintenance research

Analyze where potholes and visible road defects occur to support road maintenance research, infrastructure prioritization, or exploratory transportation analysis.

### Geospatial machine learning

Combine pothole observations with roads, weather, traffic, land use, neighborhoods, census information, road classifications, or other spatial datasets.

### Urban infrastructure monitoring

Build systems that retrieve or monitor visible changes to roads and other public infrastructure over time.

### Geographic AI

Use physical-world observations as geographic context for AI agents, multimodal systems, spatial reasoning models, and mapping applications.

## Dataset schema

The downloadable Parquet and CSV files use a normalized schema.

Source-specific attributes are preserved inside the JSON-encoded `properties` field rather than being expanded into many source-dependent top-level columns.

The GeoJSON release exposes the same attributes alongside geographic geometry.

| Column         | Type    | Description                                                                            |
| -------------- | ------- | -------------------------------------------------------------------------------------- |
| `id`           | string  | Stable identifier for the exported observation.                                        |
| `image_id`     | string  | Identifier associated with the optional image archive; null when no image was sampled. |
| `sample`       | boolean | Whether the observation was selected for the optional image archive sample.            |
| `latitude`     | float64 | Latitude in decimal degrees using WGS 84 when available.                               |
| `longitude`    | float64 | Longitude in decimal degrees using WGS 84 when available.                              |
| `geometry_wkb` | binary  | Geographic geometry encoded as Well-Known Binary for spatial workflows.                |
| `media_url`    | string  | URL for associated street-level imagery or other visual media when available.          |
| `attribution`  | string  | Attribution information preserved from the contributing source.                        |
| `source`       | string  | Source or provider associated with the observation.                                    |
| `properties`   | string  | JSON-encoded object containing source-specific attributes and retrieval metadata.      |

## Source-specific properties

The `properties` field preserves additional metadata associated with each pothole observation.

The exact fields vary by source and record.

| Property            | Description                                                                         |
| ------------------- | ----------------------------------------------------------------------------------- |
| `address`           | Address or geographic location text associated with the observation when available. |
| `captured_at`       | Timestamp of the imagery or geographic observation.                                 |
| `source_url`        | URL associated with the original source observation when supplied.                  |
| `owner / publisher` | Original owner, contributor, or publisher when supplied.                            |
| `year`              | Year associated with the observation.                                               |
| `month`             | Month associated with the observation when available.                               |
| `season`            | Seasonal context associated with or derived from the observation.                   |
| `time_of_day`       | Time-of-day context when available.                                                 |

Not every property is populated for every observation.

## Download the pothole dataset

The dataset is available in multiple formats for geospatial, machine-learning, and data-analysis workflows.

* [Parquet](https://huggingface.co/datasets/embedearth/pothole/resolve/main/data.parquet)
* [GeoJSON](https://huggingface.co/datasets/embedearth/pothole/resolve/main/data.geojson)
* [CSV](https://huggingface.co/datasets/embedearth/pothole/resolve/main/data.csv)

### Parquet

Recommended for analytics, Python workflows, DuckDB, data science, and large-scale processing.

### GeoJSON

Recommended for GIS software, spatial databases, MapLibre, Leaflet, deck.gl, web maps, and other geographic applications.

### CSV

Recommended for simple tabular analysis, spreadsheets, and general-purpose data workflows.

## Build applications with EmbedEarth

### Geographic Search

**[EmbedEarth Search](https://www.embed.earth/docs/sdk/search)** lets developers query published geographic features across areas and receive the results as geographic data.

Search can be constrained using geographic areas, coordinates, bounding boxes, properties, columns, and result limits.

Use it to power:

* geographic search engines
* mapping applications
* infrastructure analysis
* AI agents
* location intelligence
* spatial research
* monitoring systems
* geospatial data pipelines

### JavaScript and TypeScript SDK

EmbedEarth provides an SDK for building spatial search, routing, compute, mapping, and data workflows directly into applications.

**[View the EmbedEarth documentation](https://www.embed.earth/docs)**

### API

Use EmbedEarth programmatically from applications, backend services, AI systems, and data pipelines.

The platform includes geographic capabilities for Search, Compute, Routing, Geocoding, Maps, and Data.

### CLI

Use EmbedEarth directly from the terminal for geographic search, local data workflows, routing, compute, and spatial development.

### MCP for AI agents

EmbedEarth can expose geographic capabilities to MCP-compatible AI agents.

This allows agents to work with dedicated geographic infrastructure instead of relying exclusively on general model knowledge when answering questions about places and the physical world.

## Example applications

### Road condition mapping

Map pothole observations and visible pavement damage across cities, neighborhoods, transportation corridors, or other geographic regions.

### Pothole detection models

Use the dataset as part of experimentation or evaluation for visual systems designed to recognize potholes and other road-surface defects.

### Infrastructure inspection

Combine street-level observations with roads and other infrastructure data to identify areas that may warrant closer inspection.

### Road maintenance prioritization

Analyze the geographic distribution of pothole observations alongside traffic, road class, population, weather, or municipal data.

### Smart city applications

Build applications that combine road conditions with service requests, traffic accidents, permits, construction activity, and other urban data.

### Physical-world search

Use natural-language or feature-based retrieval to discover conditions and objects that exist at physical locations.

## Methodology

Observations were retrieved through semantic geographic search around the query:

**“Pothole”**

Matching may incorporate available:

* street-level visual context
* textual descriptions
* source metadata
* temporal metadata
* geographic context

The dataset prioritizes relevant, discoverable pothole observations rather than exhaustive geographic coverage.

Inclusion in the dataset indicates semantic relevance to the pothole query. It does not necessarily mean the observation has been independently inspected or formally classified as a pothole by a transportation authority or civil engineer.

## Data sources and attribution

This dataset was prepared and distributed by **[EmbedEarth](https://embed.earth)**.

Original or contributing sources include:

* [Mapillary](https://www.mapillary.com/app/)
* [iNaturalist](https://www.inaturalist.org/)

Where available, record-level source information, attribution, and source URLs are preserved in the exported dataset.

Users accessing or using individual media assets should retain applicable attribution and comply with source-specific terms.

## License

### Dataset compilation

The **EmbedEarth dataset compilation, organization, selection, and EmbedEarth-created metadata** in this repository are licensed under the **Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0)**.

You may:

* share and redistribute the licensed dataset material
* remix, transform, and build upon it
* use the licensed material for commercial or non-commercial purposes

Subject to the following conditions:

* **Attribution** — provide appropriate credit to EmbedEarth and, where applicable, the original source or contributor.
* **ShareAlike** — adaptations of material covered by this license must be distributed under CC BY-SA 4.0 or a compatible license.

Suggested attribution:

> EmbedEarth, “Pothole & Road Damage Dataset — 2023–2025,” https://embed.earth, licensed under CC BY-SA 4.0.

### Third-party content

The CC BY-SA 4.0 license applied to the EmbedEarth dataset compilation **does not replace, expand, or override licenses attached to third-party content**.

Images, videos, observations, source metadata, links, and other third-party materials referenced by this dataset may remain subject to separate licenses, attribution requirements, or usage restrictions established by their original creators or source platforms.

Users are responsible for complying with the applicable terms of any underlying third-party content they access or use.

Where available, attribution and source information are preserved in the exported records.

## Limitations

This is a **discovery-oriented pothole dataset**, not an authoritative or comprehensive road-condition inventory.

Geographic coverage depends on the availability of source observations and imagery. Coverage may vary significantly between cities, regions, and countries.

The dataset can contain:

* incomplete metadata
* uneven geographic coverage
* duplicate or nearby observations
* ambiguous road defects
* false-positive semantic matches
* observations whose road condition has since changed

Applications requiring authoritative road-condition assessments should independently validate relevant observations.

---

## Need more than a static pothole dataset?

This Hugging Face release contains **10,811 pothole observations**, but potholes are only one physical-world feature available through EmbedEarth.

**[EmbedEarth](https://embed.earth)** provides access to **millions more geographic features and large collections of geolocated images and videos**.

Search roads, infrastructure, places, buildings, community issues, natural features, physical conditions, and other real-world observations programmatically using the EmbedEarth API, SDK, CLI, and MCP.

**[Explore EmbedEarth →](https://embed.earth)**

**[Read the developer documentation →](https://www.embed.earth/docs)**

**[Search geographic features with the SDK →](https://www.embed.earth/docs/sdk/search)**