Dataset Card for muscari-data

muscari-data is the companion dataset for MuScaRi (Multi-Scale species Richness estimation, also named after the Muscari genus of perennial bulbous plants), as proposed in the paper Multi-scale species richness estimation with deep learning by Boussange et al.

It bundles ~350k anonymized European vegetation plots (EVA), 184 independent regional plant inventories (GIFT), and gridded environmental predictors, enabling development and evaluation of methods that estimate species richness and related diversity metrics from opportunistic small-scale ecological surveys.

Species names are fully anonymized; it is not possible to reconstruct the original taxonomy from the released files alone.

• Curated by: Victor Boussange (WSL / ETH Zürich) and co-authors
• Repository: https://github.com/vboussange/MuScaRi
• Paper: Multi-scale species richness estimation with deep learning
• Pretrained model: vboussange/muscari
• Demo:

Dataset Structure

The dataset bundles three interoperable components:

Component	Description	Format
EVA	Anonymized presence-absence matrix derived from the European Vegetation Archive (EVA). Each row is a vegetation plot; each column is an anonymized plant species.	Parquet
GIFT	Anonymized presence-absence matrix derived from the Global Inventory of Floras and Traits (GIFT). Each row is a regional flora (polygon); each column is an anonymized plant species. Used as an independent out-of-distribution evaluation set.	Parquet
Environmental Features	Gridded environmental predictors over Europe: 19 CHELSA v2 bioclimatic variables (1981–2010), EEA Digital Elevation Model (1 km), and Corine Land Cover 2018.	NetCDF

Together these components provide the inputs and ground-truth supervision to train and evaluate MuScaRi, and to generate continent-scale species-richness projections.

Dataset creation

Sources

EVA: European Vegetation Archive

The EVA component is derived from the European Vegetation Archive (Chytrý et al., 2016; version 2023-02-04, project 172), a coordinated repository of vegetation-plot records contributed by national databases. The raw dataset contains 502,724 vegetation plots; after filtering for plots with available area information, coordinate uncertainty < 1 km, land location, recording year ≥ 1972, and removing non-vascular taxa, approximately 352,000 plots covering ~8,500 distinct vascular plant species across 44 European countries are retained. Raw EVA data require a data-sharing agreement; the version released here has been:

1. Sanitized: plots with unreliable coordinates, missing metadata, or implausible species counts removed; non-vascular taxa and pre-1972 records excluded.
2. Anonymized: original taxonomic names replaced with consistent anonymized tokens shared between EVA and GIFT, so that the dataset can be distributed freely without violating EVA data-sharing restrictions.

Geographic coverage: continental Europe and Iceland (~10°W – 40°E, 34°N – 72°N).

GIFT: Global Inventory of Floras and Traits

The GIFT component is derived from GIFT (Weigelt et al., 2020), a global database of regional plant checklists and trait information. Inventories were filtered to those falling within the geographic range of the EVA plots, yielding 184 exhaustive species surveys. GIFT polygons mostly correspond to countries or administrative regions (median area ~11,700 km²) and serve as an independent out-of-distribution benchmark for evaluating total species richness under asymptotic sampling effort. Species are aligned to the EVA taxonomic namespace via the same anonymization procedure, so that anonymized species tokens are consistent between the two matrices.

Environmental Features

Source	Variables	Resolution	CRS
CHELSA v2	All available bioclimatic variables (incl. BIO1–BIO19, sfcWind, pet), 1981–2010 climatology	~1 km	EPSG:3035
EEA Digital Elevation Model (EU-DEM)	Elevation	1 km	EPSG:3035
Corine Land Cover 2018	Land-cover class (remapped to consecutive integers)	100 m → resampled to 1 km	EPSG:3035

All rasters are reprojected and resampled to a common 1 km grid in the ETRS89-LAEA projection (EPSG:3035) before upload. The MuScaRi model uses the mean and standard deviation of four variables computed within each spatial unit (mean annual temperature bio1, annual precipitation bio12, near-surface wind speed sfcWind, and potential evapotranspiration pet), together with elevation, as its environmental features.

Curation

Preprocessing and anonymization scripts are available in the MuScaRi GitHub repository under scripts/data_processing/:

• eva_preprocessing.py: sanitizes raw EVA data (coordinate checks, area filters, taxon exclusions).
• gift_preprocessing.py: sanitizes raw GIFT data.
• anonymise_gift_eva.py: harmonizes species names between EVA and GIFT, then replaces them with consistent anonymized tokens.
• Environmental feature rasters are compiled and resampled to the 1 km EPSG:3035 grid by muscari/data_processing/utils_features.py.

Data Fields

EVA / GIFT species_matrix.parquet

Column	Type	Description
geometry	POINT / POLYGON (WKB)	Plot centroid (EVA, EPSG:3035) or region polygon (GIFT, EPSG:3035)
area_m2	float32	Plot or region area in m²
<anonymised_species_name>	bool	Presence (True) / absence (False) for each species

Species column names are anonymized tokens (e.g. sp_00042) that are consistent across the EVA and GIFT matrices; the original taxonomy is not exposed.

chelsa_dem_cache.nc

Variable	Unit	Description
bio01 – bio19	varies	CHELSA v2 bioclimatic variables (1981–2010 climatology)
elevation	m	EEA Digital Elevation Model

landcover_cache.nc

Variable	Unit	Description
landcover	categorical (int16)	Corine Land Cover 2018 class code

Data Splits

Note: The EVA and GIFT matrices cannot be used directly as training samples for species richness estimation. The MuScaRi training procedure generates triplets of (sampling effort, spatial-unit features, observed species richness) by randomly placing spatial units of varying area and pooling surveys within them. This sample-generation procedure is described in detail in the paper (Methods, Section "Sample generation and spatial block cross-evaluation procedure") and implemented in scripts/data_processing/compile_sbcv_eva_samples.py and scripts/data_processing/compile_gift_samples.py.

The dataset does not ship with pre-defined train/test splits; these are constructed programmatically by the MuScaRi training pipeline using spatial block cross-validation. See the paper for details.

• Train / validation: EVA-derived triplets, partitioned into spatial folds to minimize spatial autocorrelation leakage.
• Test (in-distribution): held-out EVA-derived triplets from the fold not seen during training.
• Test (out-of-distribution): 184 GIFT regional inventories, geographically and methodologically independent of EVA, used to evaluate total species richness prediction under asymptotic sampling effort.

Quick Start

from muscari.data_processing.utils_eva import EVADataset
from muscari.data_processing.utils_gift import GIFTDataset
from muscari.data_processing.utils_features import EnvironmentalFeatureDataset

# Load vegetation-plot data (downloads automatically on first call)
eva_df  = EVADataset.from_hub()        # GeoDataFrame, EPSG:3035 point geometries
gift_df = GIFTDataset.from_hub()       # GeoDataFrame, EPSG:3035 polygon geometries

# Load environmental predictors
env_ds, lc_ds = EnvironmentalFeatureDataset.from_hub()
print(list(env_ds.data_vars))          # ['bio01', ..., 'bio19', 'elevation']

# Inspect the species matrix
species_list = eva_df.attrs['species_list']
print(f"{len(eva_df)} plots × {len(species_list)} species")

Citation

If you use this dataset, please cite both the MuScaRi paper and the underlying data sources:

@misc{boussange2025muscari,
  title   = {Multi-scale species richness estimation with deep learning},
  author  = {Victor Boussange and Bert Wuyts and Philipp Brun and
             Johanna T. Malle and Gabriele Midolo and Jeanne Portier and
             Théophile Sanchez and Niklaus E. Zimmermann and
             Irena Axmanová and Helge Bruelheide and Milan Chytrý and
             Stephan Kambach and Zdeňka Lososová and Martin Večeřa and
             Idoia Biurrun and Klaus T. Ecker and Jonathan Lenoir and
             Jens-Christian Svenning and Dirk Nikolaus Karger},
  year    = {2025},
  eprint  = {2507.06358},
  archivePrefix = {arXiv},
  primaryClass  = {q-bio.PE},
  url     = {https://arxiv.org/abs/2507.06358},
}

@article{chytry2016eva,
  title   = {European Vegetation Archive (EVA): An Integrated Database of
             European Vegetation Plots},
  author  = {Chytrý, Milan and others},
  journal = {Applied Vegetation Science},
  volume  = {19},
  number  = {1},
  pages   = {173--180},
  year    = {2016},
  doi     = {10.1111/avsc.12191},
}

@article{weigelt2020gift,
  title   = {{GIFT} -- A Global Inventory of Floras and Traits for
             macroecology and biogeography},
  author  = {Weigelt, Patrick and König, Christian and Kreft, Holger},
  journal = {Journal of Biogeography},
  volume  = {47},
  number  = {1},
  pages   = {16--43},
  year    = {2020},
  doi     = {10.1111/jbi.13623},
}

Contributions

Dataset compiled and released by Victor Boussange (WSL / ETH Zürich) and co-authors. Raw EVA data are subject to the EVA data-sharing agreement; only the anonymized derivatives are distributed here. Raw GIFT data are publicly accessible at https://gift.uni-goettingen.de.