FedCal/urbex-lost-places-worldwide

Urbex & Lost Places Worldwide — Georeferenced Dataset

A georeferenced catalogue of 61,995 abandoned / "lost" places worldwide, each with WGS84 coordinates, a normalized category, an estimated danger level, and (where available) a cover image, free-access flag and descriptive keywords.

Useful for: geospatial analysis and visualization, POI/feature extraction, clustering, "places near me" recommenders, map demos, and travel/heritage research.

⚠️ Safety & legality notice. Listed places are often hazardous and/or access-restricted structures. This dataset is for documentation and research only: it does not encourage trespassing (a criminal offence) or unauthorized exploration. Always verify ownership, structural condition and local law. Take nothing but pictures, leave nothing but footprints.

Provenance & license

This release is sourced 100% from LostFoundations (lostfoundations.org), collected from public per-place SEO pages (each exposing a schema.org/Place JSON-LD block), fetched politely (rate-limited, identified User-Agent) under the written authorization of the site operator. Place names are pseudonymized at source, and most descriptions are auto-generated from satellite/aerial imagery.

The schema also supports an osm source (OpenStreetMap via Overpass, ODbL, © OpenStreetMap contributors); no OSM rows are included in this release but the source column is kept so OSM-enriched rows can be merged later without a schema change.

source value	Origin	License	In this release
lostfoundations	LostFoundations public pages	lostfoundations-authorized (operator-authorized)	61,995 (100%)
osm	OpenStreetMap / Overpass	ODbL-1.0	0

License field: other → mixed-odbl-and-authorized. Redistribution of the LostFoundations rows is covered by the operator's authorization; if you add OSM rows you must attribute "© OpenStreetMap contributors" under ODbL.

Files

File	Format	Description
urbex-lost-places-worldwide.jsonl	JSON Lines	One place per line, full record (16 fields). Best for streaming / datasets.
urbex-lost-places-worldwide.csv	CSV	Same records, tabular. Best for pandas / spreadsheets.
urbex-lost-places-worldwide.geojson	GeoJSON FeatureCollection	Point geometry + display properties, ready for any map (maplibre/leaflet/QGIS). Includes a metadata.notice safety string.

Schema (16 columns)

column	type	description
source	string	Record origin: lostfoundations | osm.
source_id	string	Opaque original ID in the source (e.g. Mongo ObjectId, OSM node id).
source_url	string	Canonical URL of the place page in the source.
license	string	Per-record license tag (lostfoundations-authorized | ODbL).
name	string	Place name (pseudonymized by the source).
category	string	Normalized urbex category (12 classes — see below).
lat	number	Latitude, WGS84 (decimal degrees).
lon	number	Longitude, WGS84 (decimal degrees).
country	string	Country name as returned by the source (mostly German exonyms) — see Limitations.
locality	string	Partial address / town / area.
description	string	Free-text description when available (≈13% of rows).
image_url	string	Cover image URL (satellite/aerial), present on ≈100% of rows.
is_accessible_for_free	boolean	Whether the place is freely accessible (no entry fee).
danger_level	integer	Estimated hazard 0..3 (see legend).
keywords	string	Comma-separated descriptive tags when the source provides real ones (≈14% of rows).
dedup_key	string	Idempotent dedup key (lf:<id> / osm:node:<id>), unique per place.

Category distribution (12 classes)

category	count
other	53,762
residential	2,727
industrial	2,202
agricultural	1,900
military	598
transport	281
leisure	184
castle_ruins	90
infrastructure	76
medical	67
educational	55
religious	53

Why so much other? ~87% of LostFoundations places are bulk-imported "stubs" with no type descriptors (only a generic Imported keyword and a boilerplate description). other is the correct label for a geolocated point with no type signal — it is not a classifier failure. Rows that carry real keywords are categorized accurately (castle → castle_ruins, hospital → medical, farm → agricultural, station → transport, …).

Danger level legend & distribution

level	meaning	count
0	none / unknown	54,815
1	low	1,040
2	medium	5,758
3	high	382

Field completeness

field	non-empty rows	coverage
image_url	61,990	~100%
is_accessible_for_free	61,995	100%
keywords (real)	8,458	~13.6%
description	8,319	~13.4%

Geographic coverage

182 countries. Top 20 (country names as returned by the source):

country	count	country	count
Deutschland	13,330	Portugal	1,316
Italien	7,673	Finnland	1,311
Vereinigte Staaten	4,853	Kanada	1,240
Polen	4,089	Australien	952
Spanien	3,849	Rumänien	822
Frankreich	2,831	Griechenland	742
Vereinigtes Königreich	1,972	Schweiz	710
Tschechien	1,962	Kroatien	643
Österreich	1,925	Slowenien	613
(others)	…	Slowakei	607

Usage

# Hugging Face datasets
from datasets import load_dataset
ds = load_dataset("FedCal/urbex-lost-places-worldwide",
                  data_files="urbex-lost-places-worldwide.jsonl", split="train")
print(ds[0])

# pandas (CSV)
import pandas as pd
df = pd.read_csv("hf://datasets/FedCal/urbex-lost-places-worldwide/urbex-lost-places-worldwide.csv")
print(df.category.value_counts())

# Map it (GeoJSON) — any GIS / web map
import json
fc = json.load(open("urbex-lost-places-worldwide.geojson", encoding="utf-8"))
print(fc["metadata"]["count"], "places")

Limitations & ethics

• Country names are mostly German exonyms (the source localizes to German): e.g. Deutschland, Vereinigte Staaten, Italien. Normalize to ISO country codes before cross-dataset joins.
• Sparse type metadata: ~87% of rows are other. Descriptions and real keywords exist for ~13–14% of rows only.
• Pseudonymized names: place names are not real toponyms; do not treat name as a geocoding key — use lat/lon.
• Coordinates are at full precision in this release. A coordinate-fuzzing option exists upstream for high-risk sites but is disabled here by design.
• Not an invitation to trespass. See the safety notice above.

Citation

@dataset{calo_urbex_lost_places_2026,
  author    = {Calò, Federico},
  title     = {Urbex & Lost Places Worldwide — Georeferenced Dataset},
  year      = {2026},
  publisher = {Hugging Face / Kaggle},
  url       = {https://huggingface.co/datasets/FedCal/urbex-lost-places-worldwide}
}

Also on Kaggle: https://www.kaggle.com/datasets/federicocal/urbex-lost-places-worldwide. Maintained by Federico Calò — https://federicocalo.dev.