Our goal is to figure out where a particular species might be able to live on the planet! An Area of Habitat (AoH) map refines a species range polygon down to the parts of that range the species can live in. AoH maps feed global metrics such as LIFE (paper), the IUCN STAR metric and protected area planning.
An AoH needs a range polygon, the habitat classes the species uses in the IUCN habitat classification scheme, and its elevation limits. Finally, adding a global habitat map and a digital elevation model, and our aoh-calculator produces the raster.
The first three elements of the AoH come out of the primary literature, and are usually read by a trained assessor for a few species at a time. Global AoH maps therefore exist for birds and mammals and for few other groups. Under 2% of invertebrates have a Red List assessment at all.
1 Project overview
Our hypothesis is that an agentic pipeline over the published literature can recover habitat classes, elevation limits and range polygons at an accuracy assessors accept, for taxa that have no Red List assessment.
1.1 Experimental structure
The pipeline would run over the Evidence TAP PDF corpus, which is the successor to our Conservation Evidence work (AI-assisted Living Evidence Databases for Conservation Science). It holds around 10 million conservation fulltexts. For each species, the pipeline can emit one structured record per source document (exact format TBD):
habitats: [1.6 Subtropical/Tropical Moist Lowland Forest, 1.9 ...]
elevation: {min: 200, max: 1800, unit: m}
seasonality: resident
range_geom: <polygon, georeferenced from Figure 2>
as_of: 1998
provenance: [p4 para 2, Table 1 row 7, Fig 2 caption]
We first aim to target under-assessed groups such as invertebrates, amphibians, reptiles or plants. Birds and mammals already have several published AoH maps, but we could try to revalidate those too.
There are two excellent studies that have done something similar:
- Mining for species, locations, habitats and ecosystems extracts entities from invasion biology papers with GPT-4o, but mishandles ecological terminology.
- An evaluation of LLMs for Red List assessment finds models extract candidate range statements, then fail to turn them into a habitat coding.
For parsing from PDFs, newer models like Nemotron Parse 2.0 give us the reading order and bounding boxes, and NVIDIA's literature pipeline pairs it with a VLM over the figure regions. A figure yields a meaningful polygon only once georeferenced from coastlines, graticules or country outlines and then vectorised, for which the prior art is scanned map digitisation or geographic information in articles. Range maps add varying projections, hand-drawn boundaries, hatching for uncertain presence and inset panels, all of which complicate validation. The paper/figure legend often says which is which, so it's probably important to read the image and caption together for the VLM.
However, performance also matters. For a VLM, recall falls as the context window fills, so we will need just-in-time context by carrying references and fetching a page region on demand. This does have the advantage that each field can then record the documents consulted, those rejected and why, and the model behind each call.
It's also important to keep track of the dates referenced in a picture, since a range described in 1970 requires matching with a habitat map from back then. Disagreement between papers can then be interpreted as a range shift hypothesis, as distinct from an extraction bug.
2 Validation
We'll score against held-out IUCN habitat and elevation codings, over the papers those assessments already cite (we need to be careful about frontier models since they'll have a lot of this knowledge in their pretraining set already).
Once done, we can also push the extracted and official parameters through aoh-calculator and diff the resulting rasters. While a 50m elevation error can often leave AoH unchanged, a wrong habitat class can alter it by orders of magnitude. Michael Dales maintains this pipeline for both LIFE and STAR.
Finally, we'll take a carefully curated set of disagreements to IUCN assessors for resolution.