Skip to content
Version 1.0 · in discussion at ESSD

FullBasin

Based on MERIT DEM (Yamazaki et al., 2017) and MERIT Hydro (Yamazaki et al., 2019).

A global 90 m basin dataset, computed on MERIT Hydro across all 22.26 billion land pixels, and released with a DOI. Licence by layer: the upstream index (dfs_in, dfs_out) under MERIT Hydro’s terms, ODbL 1.0; every other layer CC BY 4.0.

Download from Zenodo
The idea

A drainage network is normally distributed as something you traverse. This one is distributed as something you look up. Two numbers at every pixel turn “what is upstream of here?” into an integer comparison, so the answer exists at any pixel rather than at a chosen list of outlets. Among published global hydrographic products, MERIT-FullBasin is the first distributed as a per-pixel queryable flow-tree index. The three tiers below follow from that one change.

Resolution
3 arcsec, about 90 m
Coverage
the whole MERIT Hydro land mask, without gaps
Outlets
24,587,290 — 24,577,574 river mouths on the coast and 9,716 inland sinks — every land pixel assigned to the one it drains to; 340,991 of their basins are above 1 km², and those alone cover >99.5% of the land
Layers
basin and region masks · per-pixel upstream index · 16 upstream attributes
Archive
100 files, 142.2 GB: 96 regional packages, global_rasters.zip, global_tables.zip, the code, and a README
Licence
by layer: upstream index ODbL 1.0, other layers CC BY 4.0 · published 22 May 2026
The three tiers

What is in the release

Tier 1Where does a pixel drain to?

Delineating a basin means recovering the complete catchment that drains to an outlet — every upstream pixel, bounded by the drainage divide. Tier 1 does this for the whole land surface at once. The basin-ID raster covers every land pixel, and the attribute table holds one row for each of the 24,587,290 outlets — from the Amazon down to single coastal pixels. 340,991 of them are above 1 km², and those alone account for more than 99.5% of the land. Raster and table share one addressing rule — basin_id = table_index + 1. Going from a pixel to its record is therefore an array offset: no join, no dictionary, no lookup table.

Tier 2What drains into a pixel?

Each pixel stores a depth-first interval. Whether one pixel lies upstream of another is then a single interval-containment comparison, independent of how large the catchment is — though listing the pixels the test returns still scales with how many there are. The index is derived from the D8 network and carries its terms: dfs_in and dfs_out are released under ODbL 1.0, as MERIT Hydro is. The MERIT Hydro layers themselves are not redistributed.

Tier 3What is upstream of a pixel like?

Tier 3 provides sixteen upstream attributes — eight terrain statistics and eight basin-shape metrics — for every pixel with at least 10 km² of upstream drainage area. Each shape metric needs that pixel’s own catchment divide, which is why attributes like these are normally published for a fixed list of basins. Tier 2 makes the divide fall out of two stored numbers, and that is what makes doing it at every pixel affordable.

Layer forms

The basin layers come in two forms (the vector one on this site, in Zenodo with v2): bsn, the 90 m raster in which every land pixel carries the ID of its basin, and bnd, the boundary of those basins as simplified vector lines. The upstream index and the sixteen attributes are 90 m rasters. The MERIT Hydro layers underneath are not redistributed; they come from the MERIT Hydro team under their own terms.

Walk-through

A recorded walk-through

A recorded talk on YouTube, covering how the three tiers are built and what each one is for. The talk plays here in the page; nothing loads from YouTube until you press play.

A note on the narration: the talk is Lulu Jiang’s throughout, but the voice you hear is synthetic, made with MiniMax. She has been abroad only once, on a short trip, and is not yet confident recording it herself. If you speak English well and would like to volunteer to record it, we would welcome that. Please write to [email protected]. The content will not change when the voice does.

Cost and validation

What it cost to build, and how it was checked

Fine-resolution hydrography underpins flood prediction, large-sample hydrology and machine-learning streamflow models, yet no global product delivered basin identity, upstream topology and morphometry together at pixel resolution. The obstacle is scale and connectivity at once: computing per-pixel attributes across a 75.17-billion-cell grid, 22.26 billion of them land, exceeds the memory of standard GIS software, and the naive fix — tiling — breaks the very basins being measured.

The complete pipeline ran in about 21 hours on one 13-thread node with terabyte-scale memory, processing basin-complete regions, so that no drainage basin is cut across an arbitrary raster tile. Each tier was checked on its own terms:

TierWhat was checkedAgainst what
1 · basin masksthe pixel count of every basinexact agreement with the flow-accumulation raster
2 · upstream indexthe interval index checked pixel by pixel against the D8 network it was built fromconsistent for every pixel
3 · attributescatchment perimeter, the most error-sensitive of the sixteen, on about 150,000 upstream-catchment masksthree independent perimeter implementations

Perimeter is the one benchmarked because a single misplaced boundary pixel moves it by about 4/P against 1/A for an area, with A and P the pixel counts of the catchment and its boundary. The manuscript gives the full validation.

Download FullBasin from Zenodo 100 files, 142.2 GB · licence by layer · 10.5281/zenodo.20344113 — or start from the basin and region mask and take one region at a time.

The manuscript is in open discussion at Earth System Science Data; the discussion paper is open for comment there.

Site under development — pages may change daily. The current data release is archived on Zenodo. Comments on the dataset are welcome in the open discussion at ESSD; anything else, [email protected]