Lilambd/world-earth

World Earth Surface Ledger: what the ground of every region is made of, and how it changed

🌐 日本語 · 中文 · Español · Português · Français · Deutsch · 한국어 · हिन्दी · العربية

• 日本語 — World Earth Surface Ledger：世界の地表変化データ: 各地域の地表が何でできていてどう変わったか。2001-2025年の毎年の森林消失、2000-2020年の土地被覆、1984-2021年の地表水、市街地、土壌を30 m衛星ピクセルから集計。
• 中文 — World Earth Surface Ledger（全球地表变化数据）: 每个地区的地表由什么构成、如何变化：2001-2025 年逐年森林损失、2000-2020 年土地覆盖、1984-2021 年地表水、建成区和土壤，均由 30 m 卫星像元统计。
• Español — World Earth Surface Ledger (superficie terrestre por región): De qué está hecho el suelo de cada región y cómo cambió: pérdida forestal anual 2001-2025, cobertura 2000-2020, agua 1984-2021, área construida y suelos, desde píxeles satelitales de 30 m.
• Português — World Earth Surface Ledger: florestas, uso do solo e água por região: Do que é feito o solo de cada região e como mudou: perda florestal anual 2001-2025, cobertura do solo 2000-2020, água 1984-2021, área construída e solo, em pixels de satélite de 30 m.
• Français — World Earth Surface Ledger : déforestation, occupation du sol et eau par région: La surface de chaque région et son évolution : perte forestière annuelle 2001-2025, occupation du sol 2000-2020, eaux de surface 1984-2021, bâti et sols, comptés sur des pixels satellite de 30 m.
• Deutsch — World Earth Surface Ledger: Waldverlust, Landbedeckung und Wasser je Region: Woraus der Boden jeder Region besteht und wie er sich wandelte: Waldverlust je Jahr 2001-2025, Landbedeckung 2000-2020, Oberflächenwasser 1984-2021, bebaute Fläche und Boden aus 30-m-Satellitenpixeln.
• 한국어 — World Earth Surface Ledger: 지역별 산림·토지피복·수면 변화: 30 m 위성 픽셀로 집계한 지역별 지표 변화: 2001-2025 연간 산림 손실, 토지피복 2000-2020, 수면 1984-2021, 시가지와 토양.
• हिन्दी — World Earth Surface Ledger (पृथ्वी सतह खाता): हर क्षेत्र की ज़मीन किससे बनी है और कैसे बदली: 2001-2025 हर वर्ष वन हानि, 2000-2020 भूमि आवरण, 1984-2021 सतही जल, निर्मित क्षेत्र और मिट्टी, 30 m सैटेलाइट pixels से गिने गए।
• العربية — سجل سطح الأرض العالمي (World Earth Surface Ledger): مِمَّ تتكوّن أرض كل منطقة وكيف تغيّرت: فقدان الغابات كل عام 2001-2025، والغطاء الأرضي 2000-2020، والمياه السطحية 1984-2021، والعمران والتربة، من بكسلات أقمار صناعية بدقة 30 m.

Every state, province and prefecture of the world (4,596 admin-1 regions, plus _none for land outside them), counted from 30 m satellite pixels:

• tree cover in 2000 and forest loss for every year 2001-2025, at 7 canopy thresholds (10-75%)
• land cover in 2000, 2005, 2010, 2015 and 2020 (trees, short vegetation, cropland, built-up, water, desert, snow and ice), and what turned into what
• surface water that disappeared or appeared between 1984 and 2021
• built-up area and population every 5 years from 1975 to 2025
• 10 soil properties at two depths, plus soil organic carbon stock

Nothing is generated by a model. Every number is a pixel count or area-weighted mean from the public rasters below, and every one of the 3,029 source tiles is listed with URL, size and SHA-256 in source_file. Pixel coverage is complete: 280 of 280 forest tiles, 280 of 280 land-cover tiles, 403 of 403 water tiles (coverage).

Data as of 2026-11-01 (computed 2026-11-01T03:12Z from Hansen GFC-2025-v1.13, the latest release).

Quick start

from datasets import load_dataset
loss = load_dataset("Lilambd/world-earth", "forest_loss_year", split="train")
base = load_dataset("Lilambd/world-earth", "forest_base", split="train")
# forest loss per year in one region at 30% canopy
jp = loss.filter(lambda r: r["a1"].startswith("JPN") and r["thr"] == 30)

Tables

config	rows	what
forest_base	31,388	per region × canopy threshold: tree cover 2000, total loss 2001-2025, gain 2000-2012, land, water and no-data km²
forest_loss_year	634,356	per region × threshold × year: forest loss km²
lc_group	216,234	per region × year × cover group: km²
lc_transition	210,936	per region: km² that went from group in 2000 to group in 2020
lc_hist	1,717,244	per region × year × raw GLCLU class value: km² (decode with lc_class)
water_transition	40,315	per region × JRC transition class (permanent, new, lost, seasonal...): km²
ghsl	49,948	per region × year: built-up km² and population (2025 is the GHSL nowcast, flagged)
soil	89,696	per region × property × depth: mean value, unit, covered km²
features_country / features_region	1,780 / 9,503	per-country and per-NUTS/US-region features in long form (iso3 or region_id, feature, value, year, unit, source)
admin1	4,597	the region list: ISO 3166-2, HASC, Wikidata, English, local and Japanese names, area, population, NUTS/US region link
lc_class / class_envo / envo_term	small	class legend, the cover groups mapped to ENVO environment terms, and the ENVO term list
source / source_file / coverage	small	datasets, every tile file, and tile coverage

Shared keys

a1 is the same admin-1 id as in Lilambd/world-hazard, so forest loss, warming and earthquakes join row by row. Also iso_3166_2, wikidata, adm0_a3 (iso3) and region_id (NUTS: / US:).

What the numbers say (computed, not curated)

• World forest loss (canopy ≥ 30%) rose from 133,680 km² in 2001 to a high of 297,331 km² in 2016; 2024 was 296,347 km², 2025 was 255,621 km².
• The largest losses by region: Sakha (Yakutia) 265,692 km², Pará 196,327, Mato Grosso 146,343, British Columbia 122,779.
• By share of 2000 tree cover, central Portugal leads: Aveiro 73.8%, Coimbra 72.4%, Castelo Branco 70.3%.
• Built-up area of the world went from 171,513 km² in 1975 to 486,982 km² in 2025 (×2.8) while population went from 4.03 to 8.11 billion.
• The most surface water lost since 1984: Karakalpakstan (−18,755 km² net) and Qyzylorda (−16,571), the Aral Sea basin.
• Within countries, regions that lost more forest did not grow faster in population (r = 0.009, n = 2,793). Built-up growth does track population growth (r = 0.097, p < 1e-9).

Gaps and known limits

• "Loss" is stand-replacing canopy loss in a pixel, including fire, logging and storms; it is not net deforestation. Gain is only for 2000-2012.
• Boreal regions (Russia, Canada) have large losses partly from wildfire.
• The country features file says "GFC v1.12" in its source text; the region tables were rebuilt on v1.13. Prefer forest_base and forest_loss_year.
• Some Natural Earth local names are wrong (for example Sakha shows the local name of Chukotka). Use name, a1 and wikidata as keys.
• Soil values are SoilGrids model predictions from profiles (c. 1905-2016), static, not a time series.
• Region area is counted on a 0.001° grid, so small islands may be off by a few km².

Use it in production

For AI agents: 14 more Lilambd open-data APIs (airports, regions, markets, security, travel and more), pay per call with x402 (from $0.005 USDC on Base, no signup), free sample with sample=1, also as an MCP server: https://api.lilambd.com/llms.txt

Need a per-place answer ("how much forest did this district lose, and when?"), a field that is not here, or a yearly feed? Write to hello@lilambd.com.

Related: Lilambd/world-hazard (earthquakes and warming per region), Lilambd/world-signals (what the world pays attention to, daily).

Quick start: datasets, pandas, DuckDB

import pandas as pd  # hf:// paths need `pip install huggingface_hub`
BASE = "hf://datasets/Lilambd/world-earth/data/"
loss = pd.read_csv(BASE + "forest_loss_year.csv")
admin1 = pd.read_csv(BASE + "admin1.csv", keep_default_na=False, na_values=[""])
world = loss[loss["thr"] == 30].groupby("year")["loss_km2"].sum()
print(world.tail())

ghsl = pd.read_csv(BASE + "ghsl.csv")
print(ghsl.groupby("year")[["built_km2", "pop"]].sum())

-- DuckDB 0.10+ reads hf:// paths directly
-- regions with the most forest loss 2001-2025 at 30% canopy
SELECT b.country, b.name, sum(l.loss_km2) AS loss_km2
FROM 'hf://datasets/Lilambd/world-earth/data/forest_loss_year.csv' l
JOIN 'hf://datasets/Lilambd/world-earth/data/admin1.csv' b USING (a1)
WHERE l.thr = 30
GROUP BY ALL
ORDER BY loss_km2 DESC LIMIT 10;

A runnable notebook is in notebooks/quickstart.ipynb.

Use cases

• Yearly forest-loss series per state or province for ESG, supply-chain and deforestation screening.
• Land-cover change (2000-2020) and built-up growth (1975-2025) per region for urbanisation studies.
• Region features that join on a1 with world-hazard and world-ground for multi-risk models.
• Teaching remote-sensing aggregates without handling rasters.

Citation

@misc{lilambd2026_world_earth,
  author       = {Lilambd},
  title        = {Lilambd World Earth Surface Ledger},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Lilambd/world-earth}},
  url          = {https://huggingface.co/datasets/Lilambd/world-earth}
}

Please also credit the original sources listed in this card.

Sources and licence

• Hansen / UMD / Google / USGS / NASA, Global Forest Change GFC-2025-v1.13 (Hansen et al. 2013, updated). CC BY 4.0.
• GLAD, University of Maryland, Global Land Cover and Land Use Change 2000-2020 v2 (Potapov et al. 2022). CC BY 4.0.
• European Commission JRC, Global Surface Water v1.4 (Pekel et al. 2016). CC BY 4.0.
• European Commission JRC, GHSL GHS-BUILT-S and GHS-POP R2023A. CC BY 4.0.
• ISRIC SoilGrids 2.0 (Poggio et al. 2021). CC BY 4.0.
• Environment Ontology (ENVO) via EMBL-EBI OLS4. CC0 1.0.
• Natural Earth admin-1 10m. Public domain.

The derived tables are released under CC BY 4.0. Please cite "Lilambd World Earth Surface Ledger" and the original authors above.

Maintained by Lilambd. Questions: hello@lilambd.com

Changelog

• 2026-11-03 — Two NUTS links corrected in features_region and admin1. In features_region, the ids NUTS:BG412 and NUTS:HU213 each had 13 features with two conflicting values, 26 conflicting rows in all. The cause was city units linked to the NUTS 3 region that surrounds them. For Bulgaria, the rows computed for Sofia city (Grad Sofiya, BG-22) are now under their correct NUTS 2021 code, NUTS:BG411 (Sofia (stolitsa)). NUTS:BG412 keeps the rows for Sofia Province (BG-23). For Hungary, the rows computed for Veszprém urban county (city, HU-VM, 227 km²) were removed, because NUTS 3 has no unit for the city alone. NUTS:HU213 keeps the rows for Veszprém County (HU-VE). features_region goes from 9,516 to 9,503 rows and from 732 to 733 region ids, and each (region_id, feature) now appears once. In admin1, the NUTS link and the OECD NUTS 3 population (2025) were moved from the city rows to the province and county rows they belong to: Sofia Province gets 224,701 and BG412, Veszprém County gets 333,345 and HU213. Grad Sofiya now links to NUTS:BG411 with population left blank, and Veszprém (city) has no link or population. admin1 keeps 4,597 rows.