Predicting EUNIS habitats across Europe, pixel by pixel

What is being predicted?

EUNIS is a hierarchical classification of European habitats widely used in biodiversity assessments and planning. Our system predicts the most likely EUNIS type for each pixel. We first infer detailed Level 3 habitats and then derive Level 1 and Level 2 from the official hierarchy, enabling consistent exploration across scales. In total, we produced maps for 200+ Level 3 habitats; the maps you see aggregates them to Levels 1 and 2 for readability.

How we made it (in a nutshell)

We built the maps through a cascading pipeline that couples species distribution modeling with habitat classification — think of it as moving from “who grows where?” to “which habitat does this assemblage indicate?”

1. Predict plant species across Europe

• Data: We assembled GeoPlant, a European‑scale dataset combining 5 million presence‑only observations (e.g., citizen science) with 90,000 presence‑absence vegetation plots from expert surveys. For each 50×50 m cell, we paired these observations with Sentinel‑2 and Landsat imagery and a 20‑year climate time series. Coordinates were not used as predictors to focus the model on habitat suitability rather than location cues.

• Model: A deep species distribution model (deep‑SDM) learns from remote sensing and climate to predict the probability of occurrence for thousands of vascular plant species, capturing interspecies dependencies at high resolution. We mitigate sampling bias using a target‑group background strategy and threshold probabilities to obtain local species assemblages.
• Scale and resolution: We tiled Europe into 25×25 km blocks and predicted at 50 m resolution for year 2021, generating 5.5 billion grid cells of species predictions (≈15 TB of data).

2. Infer habitats from species assemblages

• Model: Pl@ntBERT, a domain‑adapted language model, reads the predicted species list (ordered by estimated coverage) for each pixel and assigns the most probable EUNIS habitat type. Because it operates on species co‑occurrence patterns rather than raw imagery, it complements remote sensing and is robust to certain sampling biases. We predict Level 3 directly and derive Levels 1–2 via the EUNIS hierarchy.
• Performance: On internal benchmarks, when using the top predicted species, Pl@ntBERT achieved around 76% top‑1 accuracy at Level 1, 63% at Level 2, and 45% at Level 3—reflecting the increasing granularity of the task.

How to read the map

• Each pixel shows the single most likely habitat class. Boundaries may appear sharp where the model is confident, and more mosaic‑like in heterogeneous areas—both are expected behaviors at 50 m resolution.
• Level 1 portrays broad patterns (e.g., forests vs. wetlands), while Level 2 refines these into more specific units. Switching levels helps you navigate from overview to detail without losing the link to the EUNIS hierarchy.

What this is — and what it is not

• This is a scientifically grounded, AI‑assisted synthesis of multiple data streams, made explorable via a fast web map. By coupling species predictions with species‑to‑habitat inference, it provides a coherent, continental‑scale picture of habitat patterns at unprecedented resolution.
• It is not a replacement for field surveys or official habitat inventories. Known limitations include plot‑to‑pixel scale mismatches, uneven observation density, and the inherent ambiguity of assigning a single label to mixed pixels. Nonetheless, the approach has been validated on large independent datasets and provides actionable baselines for screening, planning, and monitoring.

Under the hood

The web app is a lightweight, map‑centric interface serving tiled layers via standard WMS, so you can integrate them in other GIS tools. Behind the scenes, we cache and stream terabytes of predictions to keep interaction fluid at continental scale.

References

Mapping biodiversity at very-high resolution in Europe

In short, the EUNIS habitat map reflects “habitats inferred from plants,” pixel by pixel, across all of Europe. Start at Level 1 for the big picture; zoom into Level 2 for nuance. And when you need detail beyond Level 1 and 2 you can use the full GeoPl@ntNet interface to explore Level 3 predictions maps and species predictions maps split by country.