GeoRas2Vec dataset – from geological map images to vector units and faults

Code, model and documentation: GeoRas2Vec on GitHub · trained model: PerfectWin/georas2vec-model

512×512 tiles of geological maps with pixel-exact ground truth for map-unit instances and faults, covering 4 continents (North America, Europe, Africa, Oceania). Built to train models that turn a scanned geological map back into GIS polygons (e.g. affinity / boundary prediction + watershed).

• real tiles: USGS Geologic Map of North America (GMNA) 300 dpi rasters, with the official GMNA vector units and faults rasterised on the same grid (~423 m/px).
• synthetic tiles: the official vector maps of each region are rendered as realistic printed geological maps (CGMW/ICS chronostratigraphic colours modulated by lithology, stipple/hatch/volcanic "v"/plutonic "+" patterns, contacts (sometimes dashed), thick faults with thrust teeth or normal-fault ticks, unit labels with leader lines, sea with isobaths and ocean names, graticule, rivers, political boundaries, place names), then optionally "printed and scanned" (paper tint, uneven lighting, blur, noise, channel misregistration, JPEG). 20 % of the tiles use a random palette so a model cannot learn "colour = unit". Scale varies from ~80 to ~600 m/px.

Splits (spatial, no leakage)

Each source is cut into geographic blocks (150–600 km); every block belongs to exactly one split and a tile is kept only if it lies entirely inside blocks of its split. Real and synthetic GMNA tiles share the same blocks.

continent	source	kind	train	validation	test	total
Africa	UNESCO-Ethiopia	synthetic	560	70	70	700
Europe	GEO-LiM	synthetic	960	120	120	1200
North America	GMNA	real	1141	81	78	1300
North America	GMNA	synthetic	480	60	60	600
North America	USGS-California	synthetic	560	70	70	700
North America	USGS-Nevada	synthetic	480	60	60	600
Oceania	GA-Australia	synthetic	1200	150	150	1500

Columns

column	content
image	RGB map tile (PNG, 512×512)
instances	uint16 PNG: 0 = no mapped unit (sea, lake, outside the map), 1..N = connected components of one geological unit. Two touching polygons of the same unit form one instance (no boundary is visible between them on a map). A minimum mapping unit is applied before rendering: polygons smaller than 40 px or thinner than ~4 px (invisible under contact lines) are merged into their neighbours, so image and labels always agree.
faults	PNG {0,255}: fault traces rasterised 1 px wide (concealed/inferred faults included)
units	JSON: for each instance id → unit code, age and lithology from the source map
source, continent, kind (real/synthetic), split	provenance
px_size_m, crs, bounds	ground resolution and georeferencing of the tile
render	JSON rendering parameters of synthetic tiles (palette, scan effects…)

from datasets import load_dataset
import numpy as np
ds = load_dataset("PerfectWin/georas2vec", split="train")
x = ds[0]
rgb = np.array(x["image"]); inst = np.array(x["instances"]).astype(np.int32); faults = np.array(x["faults"]) > 0

Test tiles of every source

Sources & licences (attribution required)

region	dataset	licence
North America	Garrity, C.P., Soller, D.R., 2009, Database of the Geologic Map of North America, USGS Data Series 424	public domain (USGS)
California, Nevada	Ludington, S. et al., 2005, Preliminary integrated geologic map databases for the United States – Western States, USGS OFR 2005-1305	public domain (USGS)
Europe	Donnini, M., Marchesini, I., Zucchini, A., A new Geo-Lithological Map (Geo-LiM) for Central Europe, Zenodo, doi:10.5281/zenodo.3530257	CC BY 4.0
Africa	Geological Map of Ethiopia at 1:1,000,000 scale, Verner et al., UNESCO IHP-WINS	CC BY-SA
Oceania	Geoscience Australia, Surface geology of Australia 1:1M (dataset 149179)	CC BY 4.0

Because the Ethiopian map is CC BY-SA, the whole dataset is released under CC BY-SA 4.0. The generator code (code/build_tiles.py, also in the GitHub repository under dataset/) and its configuration (build_config.json) are included.

Versions

version	date	change
v2.0 (current)	2026-11-12	Minimum mapping unit applied to the labels (and to the rendering of synthetic tiles): polygons < 40 px or with half-width < 1.5 px are merged into their neighbours, at render and at final resolution. In v1 ~58 % of the "small" instances were rasterisation slivers or fragments hidden under contact lines, i.e. not visible in the image; they made small-polygon metrics meaningless. Same sources, same spatial splits, same seed.
v1.0	2026-11-07	first release (6,600 tiles). Still available: load_dataset("PerfectWin/georas2vec", revision="v1.0")

The repository was renamed from PerfectWin/geocarto to PerfectWin/georas2vec (old links redirect).

Trained model: PerfectWin/georas2vec-model · code: GeoRas2Vec on GitHub.