Every georeferenced occurrence record in the GBIF monthly snapshot of 1 October 2026 (3,954,221,790 records read, 3,647,551,611 with usable coordinates and not marked absent), counted on a fixed 0.5 degree grid. One row per cell: how many records, how many distinct species, which kingdoms, how old the records are, and how they were made. Cells in the land mesh with no record at all are kept as rows with zeros, so an empty area is visible as "no record", not silently missing.

Nothing is modelled or estimated; every number is a count over the snapshot. Compiled by Lilambd.

Quick start:

from datasets import load_dataset
cells = load_dataset("Lilambd/world-biodiversity-cells", "cells", split="train")

Grid

cell_id = row * 720 + col, row = floor((90 - lat) / 0.5), col = floor((lon + 180) / 0.5). 259,200 cells cover the globe; this file has the 214,398 that have a record or belong to the land mesh. lat_center / lon_center are cell centres. The same key is used by the other Lilambd world faces (markets, world mesh layer), so tables join on cell_id. country (ISO3) and admin1 come from the Lilambd world mesh layer and are filled for its 91,815 cells (in_world_mesh = true).

Files

file	rows	what
`cells.parquet`	214,398	one row per cell, columns below
`cell_decade.parquet`	2,935,839	records per cell x decade x kingdom x basis of record
`species_cells.parquet`	1,326,741	per GBIF species key: kingdom, class, records, number of cells it was recorded in
`world_totals.csv`		records per kingdom x class x basis x decade, whole world

cells columns: cell_id, row, col, lat_center, lon_center, country, admin1, in_world_mesh, records, records_open, species, species_open, species_animalia, species_plantae, species_fungi, species_birds, species_insects, species_seen_once, species_only_in_this_cell, species_in_10_cells_or_fewer, records_animalia, records_plantae, records_fungi, records_birds, records_human_observation, records_specimen_or_sample, records_machine_observation, records_2020s, records_2010s, records_before_1970, records_no_year, first_year, last_year, record_status, source, source_url, source_snapshot, fetched_at.

species counts distinct GBIF specieskey values (records identified only to genus or higher add to records but not to species).
species_only_in_this_cell: species recorded in exactly one cell worldwide (385,653 species have this). Often real narrow ranges, often just one collecting trip; read it as "recorded once in space", not as endemism.
*_open columns count only records licensed CC0 or CC BY 4.0 (3,026,672,393 records). The other columns also include CC BY-NC 4.0 records.
decade = -1 in cell_decade means the record has no year.
Every row carries source, source_url, source_snapshot (2026-10-01) and fetched_at (2026-11-02).

What the counts measure (and what they do not)

These are counts of records, which follow where people look, not only where life is.

Birds are 2,337,204,993 of the 3,647,551,611 records (64%); human observations are 3,177,079,050 (87%). Cells near birding and citizen-science hubs dominate the record counts.
1,566,100,550 records (43%) are from the 2020s; 78,201,756 are from before 1970; 112,399,737 have no year.
The richest cell by recorded species is in French Guiana (32,382 species), followed by Sydney (27,680) and southern Scandinavia, which reflects survey effort as much as diversity.
21,746 land-mesh cells have no record at all, most of them in Antarctica (16,649), Greenland (1,621), Canada (707), Russia (518) and China (196).

Compare a cell with its own neighbours and its own past (cell_decade) before comparing it with the world.

How it was built

Each of the 10,448 parquet files of the GBIF AWS open-data snapshot s3://gbif-open-data-us-east-1/occurrence/2026-10-01/occurrence.parquet/ was read for nine columns only (decimallatitude, decimallongitude, specieskey, kingdom, class, year, basisofrecord, occurrencestatus, license), about 49 GB of the 303 GB. Records were dropped when coordinates were missing, out of range, exactly (0, 0), or occurrencestatus = ABSENT. No other GBIF issue flags were applied, so some records with coordinate problems remain; treat single records in remote cells with care. The scripts are in build/ (agg.py per-file partial counts, reduce.py and build.py final tables).

Licence and citation

GBIF records carry per-dataset licences (CC0, CC BY 4.0, CC BY-NC 4.0). Because most columns include CC BY-NC records, the dataset as a whole is released under CC BY-NC 4.0. The *_open columns are counted only from CC0 and CC BY 4.0 records and may be used under CC BY 4.0, including commercially, with the citation below.

Cite the source snapshot: GBIF.org (1 October 2026) GBIF Occurrence Download https://doi.org/10.15468/dl.tup2g6

Maintained by Lilambd. Questions or another grid or face: hello@lilambd.com

Downloads last month: -

Total file size:

37 MB