Invisible link to canonical for Microformats
 

Data

Collection

Species Occurrences

The species occurrence data, including geographic coordinates of individual observations, were obtained from the Global Biodiversity Information Facility (GBIF).
As GBIF enforces a delay between successive download requests (each corresponding to a single species) we implemented a polling mechanism to automate the process.
This approach enabled continuous querying and data retrieval without manual intervention, effectively eliminating idle time and improving the overall efficiency of the data acquisition workflow.

Habitat

Information on species’ conservation status and habitat preferences was obtained from International Union for Conservation of Nature (IUCN) Red List, where the relevant data are provided in individual documents.
To facilitate large-scale data collection and avoid the need for manual downloading, we developed a custom web scraper tailored to the structure of the IUCN website.
This automation significantly streamlined the process, enabling efficient and systematic retrieval of multiple files.

Climate

Climatic data were obtained from WorldClim, which provides high-resolution global climate layers.
As the data are readily accessible through the platform, they were downloaded directly via standard web access.

Cleaning

The species occurrence records obtained from GBIF were consolidated into a single file and subjected to a data cleaning process.
The cleaning process included the following steps:

  • removal of data fields not relevant to the analysis;
  • validation of date fields;
  • removal of records with a coordinate uncertainty greater than 5.000 meters;
  • application of a geographic bounding box encompassing the European continent, and removal of any occurrences falling outside this defined spatial extent;
  • removal of species represented by fewer than 20 occurrence records.

Enrichment

To get an estimate of the species’ maximum dispersal distance, we implemented a streamlined approach inspired by retrieval-augmented generation, though without the use of embedding-based similarity search, as the queries were predefined and not user-generated.
Each query targeted a particular species, and the corresponding context was extracted from species-specific documents.
Relevant sections of text were identified and isolated using natural language processing techniques.
The query and its extracted context were then jointly submitted to a LLM, from which we retrieved the desired output.

↑